Quick summary:

Podcasts hold hours of experts explaining their own work, and once an episode ends, finding a particular claim again means scrubbing through the audio. Platform transcripts and a handful of free tools can turn any episode into text fit for research, with the studies behind the conversation a click away. Simple checks then sort the rigorous shows from the rest and trace a spoken claim back to its study.

Key takeaways:

→ Podcast claims travel by word of mouth more often than by citation. A remark heard once gets repeated far beyond the episode, and going back to confirm the wording or the evidence is rare.

→ The largest audio apps now produce transcripts on their own, and paste-in websites handle the shows they miss. With text in hand, a timestamp pins a spoken claim the way a page number pins a printed one.

→ The most rigorous shows publish the studies and books behind each episode right in the episode description. Those lists separate the shows built on evidence from the rest, and they let a listener read what the expert read.

→ The same approach works on lectures and oral histories as readily as on podcasts. Speech that becomes text becomes verifiable, as more of what people know arrives as speech every year.

Scroll to the end for a recommended podcast list.

In the summer of 1933, a Harvard classics professor named Milman Parry made his first trip to what was then Yugoslavia. He carried a question that had gone unanswered since antiquity: how could the Iliad and the Odyssey, poems far longer than anyone could memorize word for word, have been composed in an oral tradition? The mountain villages of the Balkans were still home to illiterate bards who could sing heroic epics for hours at a stretch. Parry believed their way of composing was the same one that had produced the Homeric poems; recording them meant watching the ancient method still at work.

He returned the following year with a former student named Albert Lord, and this time the recorder was a custom rig: two turntables that handed off to each other so one could keep cutting a disc while the other was swapped, letting a performance run for hours without a break. Over fifteen months the two of them gathered more than twelve thousand texts, most taken down in writing and many recorded onto aluminum discs, more than three and a half thousand of them in all. Alongside the discs were ninety-five notebooks of texts taken down by dictation. What began as spoken sound became a written archive, and that archive became the foundation for how scholars understand the oral-traditional composition of the Homeric poems.

For most of history, the only way to keep something said aloud beyond the moment of its utterance was for a person to write it down. Some speech survived that way, in court records and in sermons a listener copied out. Transcription by hand was slow work, though, so those were the exceptions, and ordinary talk left no record. What Parry grasped, before magnetic tape became commonplace or digital audio existed, was that a recording is only half of the value. The other half arrives when the recording becomes text a person can search and quote.

An archive disguised as entertainment

More than half a billion people around the world now listen to podcasts, across more than four million shows. Buried in that catalog is a kind of source material that barely existed a generation ago: experts talking about their own work, at length and in their own words. On one show a scientist spends an hour on the false starts behind a published result, the kind of detail a formal paper leaves out. Another has a historian naming the exact archives behind a claim, and elsewhere people who will never write a book sit for long interviews about what they know.

The candor is one part of the value, and the paper trail is another. A published article shows conclusions, and an interview lets us hear the same person explain how they reached them and where they still hedge. The most rigorous shows then post every study and book mentioned in the episode’s notes, so a single conversation can hand us a reading list. And because shows publish in dozens of languages, the catalog reaches experts whose work never gets translated into English at all.

Some of what’s in there exists nowhere else. Plenty of expertise never gets written down at all (for example, the tacit working knowledge of surgeons, engineers, negotiators, and builders) and a long interview can be the only record of how those people think. Timing favors the medium too, since a researcher can talk through a finding the week it publishes, months before any writing that digests it could arrive. And when an interviewer pushes past a rehearsed answer, we hear something print rarely captures: an expert responding in the moment.

The clearest proof that podcasts already work as sources is how often we quote them. “I heard on a podcast” has become a standard way for a claim to enter conversation, and a statistic from an interview gets repeated at a dinner table, then again in someone’s article. Each repetition runs on memory alone. An episode link exists to send, but it opens an hour of audio with no page number, and the claim sits somewhere inside, unmarked. So the archive is already in use as a source, just without a way back to the moment, and a claim nobody can return to is a claim nobody checks.

Newsrooms already treat podcasts as source material, since a remark a public figure makes on a show routinely becomes a headline, and enough demand exists that a small industry of transcript and search tools has grown up to serve that work. So the archive has two kinds of users: professionals who mine it with purpose-built tools, and the rest of us who quote it from memory. The gap between the two is awareness, since most of us never learn the material can be searched and cited at all.

Part of why we never learn is mechanical: a conventional search engine can’t reliably search what’s only been spoken. Google, for example, can find an episode when the topic made it into the title, the written description, or an indexed transcript, but everything else—the claim made at ninety minutes in, or the study a guest named in passing—may exist only in the audio, so no typical search box will surface those words. What a search engine sees of a podcast is often the label on the box; the contents inside stay sealed.

Librarians have a name for a similar issue. Grey literature is the term for knowledge produced outside the usual commercial or academic publishing channels: government reports, working papers, conference talks, and other material that may be valuable but poorly indexed. Podcasts don’t fit the category neatly, but millions of hours of recorded expert conversation can present the same practical problem. Reaching material like that takes method: knowing where it collects, and knowing how to pull it into a usable form. For podcasts the method runs in a fixed order, and here we’ll walk through it step by step: first turning the audio into text, then finding the right shows and episodes, then checking the claims and the people who make them, and finally citing what survives the scrutiny.

Plenty of transcripts already exist

Often the first step, audio into text, is already done for us. Several of the biggest podcast platforms now generate transcripts automatically for many episodes, and each keeps the text in a different place. Platforms may create these on their own, whether or not a show publishes a transcript itself. Here’s what each offers, with the steps for opening the text.

Spotify

Spotify can show the full text of an episode while it plays, the way a music app shows lyrics. The line being spoken is highlighted, and the text scrolls in time with the audio.

Important: We must use the official Spotify Desktop App installed on our computers. If we’re using a standard web browser, the transcript features won’t show up at all.

To find the transcript:

• Play the Episode: Open the desired podcast episode and click the green Play button.

• Open the Right Sidebar: Look at the far right edge of the screen next to the main window content. Click the arrow button facing left that says Show Now Playing view when hovering over it. This slides open a wide menu pane on the right side of the screen.

• Scroll the Sidebar Down: Move the mouse cursor into that newly opened right sidebar. Scroll down completely past the cover art, the episode description, and the Comments section.

• Locate the Read Along Panel: Look right below the About the podcast block. If a transcript is available, a section titled Read along or Transcript will appear displaying the scrolling text. Tapping any line of text automatically moves the audio playback to that exact moment.

Spotify hasn’t generated transcripts for every episode. When an episode has none, no transcript option appears, and the tools in the next section fill the gap.

Apple Podcasts

Apple generates transcripts as well, and the company says more than one hundred million episodes now carry them (across thirteen regional versions of languages).

Important: We must use the official Apple Podcasts App installed on Mac computers. While there’s a web version of Apple Podcasts, the interactive transcript feature is entirely unavailable when listening through a browser.

To find the transcript:

• Play the Episode: Open the desired podcast episode and click the Play button.

• Open the Full-Screen Player: Tap the MiniPlayer bar stretching across the bottom of the screen to slide open the full-screen playback view.

• Tap the Transcript Icon: Look at the bottom-left corner of the full-screen player screen. Tap the speech bubble icon that looks like a quotation bubble with horizontal lines inside.

  

• Navigate the Text: The transcript will open in full screen and automatically scroll in sync with the audio. Tapping any line of text automatically moves the audio playback to that exact moment.

Apple hasn’t generated transcripts for every episode. If we don’t see it, or if the icon is grayed out, that means that it’s not available. Same as for missing Spotify transcripts, there are options in the next section that can help in this kind of situation.

YouTube

Many podcasts also appear on YouTube as video. When a video has captions, YouTube provides an interactive transcript text block that’s useful for locating key terms.

To open the transcript:

• Expand the Description: Locate the gray description section box directly underneath the video player screen. Click the ...more button to expand the text box completely.

  

• Open the Transcript Window: Scroll all the way to the bottom of the items inside that description section. Click the button labeled Show transcript. A dedicated transcript window will immediately slide open on the right side of the video player screen.

• Search for Keywords: Move the cursor inside that right sidebar text block. Press Command + F on a Mac, or Control + F on Windows, to search for a specific name, phrase, or term.

• Navigate the Video: Click any timestamped line of text inside the scrolling layout to automatically move the video playback to that exact moment.

YouTube can also translate the captions displayed during playback:

• Open Player Settings: Hover over the video player screen and click the Settings gear icon located on the bottom toolbar.

• Select Translation: Click Subtitles/CC, click Auto-translate, and select the desired language from the list.

When we turn on auto-translate, the new language only shows up as subtitles on the video screen, while the transcript text box on the right side of the screen stays in the original language. This means we can’t search or copy the translated words from that sidebar. Computer translations also frequently make mistakes with names and lose subtle meanings. While this feature helps us understand the general conversation, we shouldn’t rely on it for exact quotes.

For transcripts that aren’t readily available

If an episode lacks a native transcript, no panel or button will show up on its page. We can easily fix this by generating a text file directly from the audio using an external tool. These tools come in two forms: websites where we paste the episode link, or software programs that run locally on our own computers.

Paste-the-link generators

These are websites that create the transcript for us. Paste in the link to an episode, and the site returns the full text, broken into timestamps.

• Podtyper works from a public Apple Podcasts, Spotify, or YouTube episode link.

• TranscribeTube works from a YouTube link or an uploaded audio or video file, and it can label who’s speaking (which helps on a show with several guests).

Unlike the read-along panels inside Spotify and Apple, these services return a file we can save and mark up.

Whisper, for private or bulk transcription

Whisper is an open-source speech-recognition model released by OpenAI. It can run on our own computers, so the audio stays local on the machine. The practical way to use it is through an app built around it, such as MacWhisper: open the app, drop in an audio file, and the transcript appears.

For long-running podcasts, high volume is the best reason for its use: it can process a show’s back catalog in batches. It also handles audio that was never published at all (like our own interview tapes or a family’s oral history). When we use a local transcription model, those files stay on our own devices. (And to clarify, it’s unrelated to Wispr, the dictation app with the similar name.)

Looking beyond Spotify, Apple, and YouTube

Those three platforms have the easiest transcripts to reach, though podcasts play in dozens of other apps as well, with many drawing on the same public feeds. Wherever we listen, a more difficult question could be sitting underneath: which shows should we be following at all? A separate set of sites exists for answering that.

Where to find shows in the first place

These sites help us browse and search the shows themselves, by topic and, in some cases, by guest.

• Podchaser catalogs shows and their guests, the way a film database catalogs movies and actors, so searching a researcher’s name can turn up shows they have appeared on.

• Goodpods surfaces shows that other listeners recommend.

• The New Books Network runs episode after episode of scholars interviewing scholars about their new academic books, sorted by field.

• Podcast Index is an open podcast directory and database that makes public podcast feeds searchable.

The charts inside the big platforms rank shows largely by popularity. But popularity can be built with marketing, and offer nothing about whether a show’s claims and sources hold up under scrutiny. These directories offer other ways in: topic, guest, category, description, ratings, reviews, and editorial or listener recommendations. That makes it easier to find a small, rigorous show that might otherwise be buried beneath a heavily promoted one.

How to search every show at once

A directory surfaces shows to follow, and sometimes the question is narrower: which episode, anywhere, discussed a topic or quoted a researcher. Often those words can make it into an episode’s title or description, and a regular search, or Listen Notes below, will turn them up. But when they were only ever spoken over the course of the episode, transcript search engines take over.

Listen Notes

Listen Notes is a search engine for podcasts, the way Google is a search engine for the web. It indexes millions of public podcasts using the feeds shows themselves publish. Type a phrase or a name and it returns matching shows and episodes, searched across titles, written descriptions, show notes, and, where available, transcripts. It works like a card catalog for podcasting at large, the first place to look when the question is which show covered a subject at all.

Tapesearch

Tapesearch searches podcast transcripts. It has generated text for a large catalog of episodes, and typing a term returns matching passages, each with a timestamp that takes us to the moment in the audio. Exact phrases can go inside quotation marks, and search terms can be combined. For research that starts from a term and needs to find where it came up, this is the tool.

Podscan

Podscan is Google Alerts for podcasts. Give it a name or a topic, and it searches its transcript database and monitors new episodes as they publish, flagging mentions whether or not they ever reach a title or description. The other tools here help search what already exists; Podscan also keeps watch going forward, which is how a researcher tracks a topic over time and how anyone finds out they’ve been mentioned on a show.

Snipd

Snipd is built around saving and searching the parts of episodes we care about. As we listen, it lets us clip a moment, and it saves that clip with a transcript, timestamp, and summary in a personal library we can search later. Over time this becomes a searchable record of the passages we marked across everything we’ve heard.

Show notes are the works cited

Everything so far ends with words on a screen, an episode found and its transcript in hand, and words alone only prove what someone said. Checking where the words came from (that is, the provenance of these ideas) starts in the same place on every episode: the show notes, or the episode description that the show itself wrote. On most shows it carries a summary and some promotional links. The most rigorous shows use it for something else entirely: a list of the studies and books the conversation touched, with links straight to the original sources.

This is where a podcast stops being testimony and starts becoming scholarship we can verify. An expert’s spoken claim asks for our trust, and a works-cited list lets us check before we give it. With this, we can open the same papers the expert read and see whether the claim survives the trip back to its source. Whenever an episode raises a claim we plan to repeat or rely on, having these sources available allows us to read what the speakers themselves were reading and come to our own critical conclusions.

Coverage varies from show to show, and that variation can tell us something about a show’s reporting and editorial practice. Science shows tied to institutions, and interview shows that take sourcing seriously, may publish full lists of studies and books. Peter Attia’s medical show often links studies and other resources discussed in an episode, and Ezra Klein’s interviews commonly include books mentioned in the conversation and the guest’s recommendations. To find the notes, open the episode’s page in any app and expand the description. When the in-app version looks cut short, the show’s own website may carry a fuller version under the episode’s title. An episode with nothing there beyond links to sponsors has answered a question too: how much of its source trail it’s made available.

How to verify a show

The show behind an episode deserves the same scrutiny as the episode itself. Podcasting has no universal editor, reviewer, or publisher standing between a host and our ears. Anyone can start a show, and sounding authoritative is a production choice, open to anyone with a microphone. So the question of who’s behind a show comes before anything the show says.

Three quick checks reveal a great deal about a podcast:

• Search the host’s name: Relevant training, an institutional affiliation, or a career in the subject may surface quickly, along with interviews, reporting, criticism, and other evidence of how the host is regarded.

• Open an episode’s notes: A rigorous show may link its sources there; a show with nothing but sponsor links leaves its evidentiary trail harder to inspect.

• Scan the guest list: Working researchers and practitioners, whose names and work can be independently confirmed, are a useful sign that a show draws on real expertise.

There’s another shortcut for the moments when a host or guest’s name is already in hand. We often arrive at a topic from somewhere: an article that quoted an expert, or a study with an author. Podchaser can turn that name into a list of shows the person has hosted or appeared on. One search opens a shelf of candidates, each tied to someone whose work sent us looking.

Two deeper marks distinguish the strongest shows: a named institution or a stated corrections practice standing behind the work, and openness about the show’s own point of view on controversial ground. Both tend to show up on a show’s About page, as one more place to look. And production polish counts for nothing on its own: a show can be popular and beautifully produced and still cherry-pick the studies that fit its story.

How to verify what we hear

A vetted show narrows the risk, but each claim still calls for its own check because speech and print run on different standards. In conversation, an expert may speculate and simplify in ways they would rarely put in print (since talk often goes out with no peer review and less editorial scrutiny). A remark on a podcast reads best as a lead: dependable evidence of what the expert said and a pointer to published work, where a claim can be checked before we repeat it as true.

Suppose an episode claims that 80% of adults with ADHD dread replying to text messages. Before that number travels further, run it through the same checks a statistic in an article would get:

• Open the show notes and look for the study behind the number. If it’s linked, read what was measured and in how many people, since a statistic often changes shape between the paper and the retelling.

• If the notes name no source, search the claim itself: the key terms and the number, in Google Scholar or a regular search, to see whether any study produces that figure.

• Check the speaker too. A search on the expert’s name and field can turn up their training and what they have published, which helps distinguish a researcher describing their own data from someone repeating something heard elsewhere.

• A claim that traces to no study and no named data is an opinion, a hypothesis, or an unsupported assertion. Repeat it as such, or drop it.

Two cautions apply to the transcript itself. Automatic transcription mishears exactly what research leans on—proper names and figures—so before quoting anything word for word, open the audio at the timestamp and confirm the words by listening. And an contentious clip calls for the full episode, since audio can be cut in ways that reverse a speaker’s meaning, and synthetic voices are now convincing enough that a clean recording is no longer proof a person said the words. Finding the moment inside the full, original episode answers both concerns.

A long interview also offers something no pulled quote can carry. Across an hour we hear how an expert reasons, which evidence moves them and where they still hesitate. Lift one sentence out and that context drops away, which is how a hedged, qualified statement turns into a confident-sounding claim it never was. So the transcript earns its keep twice: once as a source of quotes and once as something to read in full, for the reasoning around them.

The form itself also works on us in a way print simply can’t. Hours of a voice in our ears can build the feeling of knowing the speaker; we may extend to familiar voices trust we would withhold from an unfamiliar claim on the page. A claim from a familiar voice can slip past the scrutiny we would give the same words from a stranger. The fix here is running the claim through the same checks anyway, warm familiar voice or not.

And one more bias sits underneath the medium itself. The experts who show up on podcasts are a self-selected group: the ones willing to sit for an interview and are at ease behind a microphone. Researchers who never record anything may vastly outnumber those who do, so the voices that reach us are one slice of a field, not necessarily a representative one. That gap is easy to forget when a well-argued episode starts to sound like the consensus of an entire discipline.

How to cite a podcast

After the episode is found and the claims are checked, one question is left for anyone writing a paper or a report: whether a podcast can go in the citations. It can! Citation systems have long covered recorded speech, and the style guides that govern academic writing include formats for podcast episodes. A podcast is fully legitimate as a primary source when it’s cited as evidence of what a person said, much as an interview or a speech is cited.

For a claim, though, the episode is often a conversation about the evidence. The study a guest discussed is the citation for the finding, and the episode is the trail that led there. The rest of us might not need to keep citation lists, but the habit still pays off in a smaller way: sending the timestamp and a link along with a claim lets the other person hear the moment for themselves.

A workable general citation names the creator or host, the episode title in quotation marks, the show title in italics, the publisher, the publication date, and a link to the episode. Add a timestamp when you are quoting or directing a reader to a particular claim. Include a retrieval date when a required citation style calls for one, or when the episode is unusually likely to change or disappear.

• Pattern: Host or creator, “Episode Title,” Show Name, publisher, publication date, URL; timestamp for the quoted passage.

• Example: Nature Podcast, “A black hole’s spin could finally be measured by a high-speed star,” Nature Podcast, Nature, 19 August 2026, [episode link]; 00:46.

Anyone who reads that citation can open the episode, move to 00:46, and hear the claim being made, much as a page number sends a reader to a passage in a book.

Listening like a researcher

With a transcript open and the sources in hand, an episode becomes something we can return to and question. A claim that catches our attention in minute forty no longer has to survive as a half-remembered remark. We can find the line again and follow where it points. The standards we already use on writing can now apply to what we hear.

A good podcast can do more than deliver an answer. It can lead us to a researcher’s work, a book, a paper, an archive, or a disagreement that the conversation only had time to touch. An interview may be the first place we encounter a question, not the final authority on it. Once the words are searchable, the episode becomes a way into a larger body of work. That possibility changes the listener’s role. We can admire an argument and still look at its evidence. We can take an expert seriously without accepting a spoken claim before we’ve seen what supports it. When the source holds up, we know where it came from. When it doesn’t, the search has still shown us something about the limits of the claim.

The tools that make this possible once belonged largely to newsrooms and research institutions. Transcripts, searchable audio, and episode notes have made more of that work available to anyone willing to use them. A listener who opens the notes and follows a source also changes the incentives around the medium. Shows have a reason to document their claims when listeners expect a trail they can inspect.

Recorded talk has always carried knowledge that doesn’t appear elsewhere: the texture of an interview, the hesitation before a conclusion, the working memory of a field, and the account of someone who may never write a book. What’s changed is our ability to return to it. What lectures, oral histories, conference sessions, interviews, and the meetings workplaces save for later can increasingly be searched, read, verified, and cited. A transcript leaves the listener with more than a voice to remember: it leaves the words where they can answer for themselves.

You might also like:

• The Atlas of Learning: A librarian’s map of where knowledge lives, and how to find our way through the research world.

• YouTube is a Research Library. Here’s How to Search It Like One.: 30 YouTube tools to find exactly what you need, and the verification framework underneath all of them.

• What “Do Your Own Research” Actually Means: A step-by-step methodology for evaluating any claim and exploring any subject.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides.

Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

The recommended shelf

For anyone who wants somewhere to begin, these are shows that offer strong reporting, knowledgeable hosts or guests, and a visible trail back to sources. They run in several languages and reach well past the English-speaking world. Many publish on YouTube or post useful notes and links, so the transcript and search methods above will often apply. It’s a starting point for reputable podcasts out there, not the final say on which ones to trust.

Science

• The Nature Podcast (English)
  Produced by the journal Nature, with reporting tied to newly published research and links to papers behind many stories. It runs as a brisk news briefing, and a single episode can lead to a study behind a recent finding.

• Science Vs (English)
  Wendy Zukerman takes controversial questions through the available research, with a comic touch and source lists in the notes.

• Ologies (English)
  Alie Ward interviews specialists from a different field in each episode, warm and quick, with further reading and source links in the notes.

• Coffee Break: Señal y Ruido (Spanish)
  A weekly Spanish science roundtable, loose and argumentative, that reads recent papers and reports back. The show describes itself as separating sound science from noise.

• Naruhodo (Portuguese)
  A Brazilian show that takes questions from science and everyday life through evidence and explanation. Verify the relevant episode notes when using it as a source.

• The Science Show (English)
  ABC Radio National’s long-running science program, presented by Robyn Williams and Belinda Smith. It brings researchers into conversations that range widely across contemporary science and scientific debate.

History

• The Rest Is History (English)

  Historians Tom Holland and Dominic Sandbrook talk a subject through across several episodes at a time, from antiquity to the twentieth century. The archive runs past eight hundred episodes, and the British Academy gave the show its President’s Medal (a first for a podcast).

• Empire (English)

  Historian William Dalrymple and journalist Anita Anand cover the world’s empires, from the Mughals and the Ottomans to the British. Academic guests come in for the ground each knows best, and the questions stay on how empire made the present.

• Le Cours de l’histoire (French)
  France Culture’s daily history program, presented by Xavier Mauduit, with historians discussing their research and recent books.

• Geschichten aus der Geschichte (German)
  Two historians take turns walking each other through a story from the past. It’s a weekly, source-conscious show with a large back catalog.

• Il podcast di Alessandro Barbero (Italian)
  These unofficial feeds gather public lectures by the medieval historian Alessandro Barbero into an enormous history archive. The scholarship and voice are Barbero’s, though the selection, framing, and feed management belong to unofficial publishers.

• Fall of Civilizations (English)
  Paul Cooper’s long, atmospheric episodes examine societies in crisis and collapse, global in scope and accompanied by source lists.

• Revolutions (English)
  Mike Duncan’s narrative history follows the major revolutions from the English Revolution through the Russian Revolution. Its ten-season original run remains a patient, evenhanded course in political upheaval.

• COTEN RADIO (Japanese)
  A Japanese history podcast produced by COTEN, a company that organizes historical research and turns it into long episodes on figures, events, and civilizations.

• The Ancients (English)

  Tristan Hughes interviews a different historian or archaeologist each episode on one theme from antiquity, twice a week. The guests are working academics, and the notes link their books and the related episodes.

• The History of Africa Podcast (English)
  A narrative history of African empires, kingdoms, and societies that takes a deliberately non-Eurocentric approach.

Big ideas and long-form interviews

• Sean Carroll’s Mindscape (English)
  The physicist Sean Carroll interviews people working across science, philosophy, and public life. Conversations can become technical, and their value often lies in hearing specialists pursue an idea at length.

• Dwarkesh Podcast (English)
  Dwarkesh Patel conducts unusually long, closely prepared interviews on artificial intelligence and frontier science. Use each episode as an opening into the guest’s work, not as a substitute for it.

• Conversations with Tyler (English)
  Tyler Cowen’s fast interviews range across economics, culture, technology, and public life. He often approaches economic questions from a classical-liberal perspective, a useful lens to keep in view.

• The Seen and the Unseen (English)
  Amit Varma’s extended interviews move through India’s economy, politics, history, and culture. The show's notes can be unusually extensive, and Varma brings a declared free-market sensibility to the economics.

Share Card Catalog

This week:

• OpenAI now places teenagers in a stricter version of ChatGPT, guessing each user’s age from the way they use the account.

• The private conversations people have with chatbots are surfacing as evidence in courtrooms and government files alike.

• Google’s new phone can turn American Sign Language into written English, similar to how voice dictation translates into text.

OpenAI builds a teen ChatGPT

What happened

On August 18, OpenAI introduced ChatGPT for Teens, a more guarded version of its chatbot, with extra safety limits switched on by default for anyone OpenAI takes to be under 18. There’s no separate app to download. The same ChatGPT simply tightens its rules when OpenAI’s system decides a user is a minor, and teenagers who state an age between 13 and 17 are placed in the teen version automatically.

Behind the feature sits a system OpenAI calls age prediction, which estimates how old a user is from a mix of behavioral and account-level signals, rather than relying only on the birthday entered at signup. When the system can’t tell, it errs young and applies the under-18 limits, and an adult wrongly sorted into the teen version can verify their age to get the full version back. Inside the teen version, the chatbot backs away from content OpenAI counts as unsafe or developmentally inappropriate for minors, including graphic violence, sexual or romantic roleplay, self-harm, dangerous activities, and material promoting unhealthy dieting or body shaming.

Parents can tether their own account to a teen’s and, from there, set Quiet Hours, manage selected settings, and switch off features such as memory. What they can’t do is read the conversations. If the system senses a teen in acute distress, it may alert the parent or guardian; OpenAI says such notifications are limited to high-risk situations and share only the information needed to support the teen’s safety. These protections arrive years after teenagers began using ChatGPT, and OpenAI says they draw on developmental science and guidance from outside experts.

Why it matters

OpenAI’s teen version marks a change in what a chatbot is expected to know about the person using it. The company is not merely offering a safer setting for young people; it’s building a system that sorts users into different versions of the service according to an inferred trait. That makes age less a fact a person supplies than a judgment the platform makes.

OpenAI isn’t acting alone. In recent years, regulators across Europe and the United States have been writing rules that push online platforms to establish how old their users are, and OpenAI is among the first to build that work into a chatbot used by hundreds of millions of people. If it works at scale, other companies may follow: a way to sort children from adults without making everyone show an ID.

The immediate question is whether the system protects teenagers. The longer one is what becomes normal once platforms begin drawing conclusions about their users without being directly told: how much they can infer, what those inferences can decide, and how a person can challenge a wrong one.

What it means for us

The age-guessing doesn’t stop at teenagers. When OpenAI’s guess about our own age misses, its fix is to ask for a selfie or, in some cases, a government ID. That’s the bargain now settling over the services we use every day: accept the software’s read on who we are, or hand over more of ourselves to prove that read wrong.

For anyone raising a teenager, the change is more immediate. The version their child sees now hangs on OpenAI’s guess about the child’s age, and that guess is imperfect. A teenager can enter a false age, even if OpenAI says its system is designed to look beyond what the account holder declares. That’s reason enough to switch the parental controls on rather than trust the guess alone. For the rest of us, the lesson is to notice what a tool assumes about us by default, and what it needs to collect to make those assumptions. Those decisions are often made long before we know to question them.

Bottom Line:

The age check is the visible part. Underneath it, OpenAI has built an engine that studies behavior closely enough to make a judgment about a user’s profile. Once a company can sort people this way in the name of child safety, it’s established a broader habit: using what people do on a platform to decide what they should be allowed to see there.

Age is only one trait a platform might try to infer. Other systems already claim to read mood, interests, or political leanings from behavior, often to decide what to show a person, when to show it, and how best to hold their attention. Child protection is the rare aim almost no one will argue against, which makes it a powerful justification for building this kind of machinery. The harder question is where the boundary lies: what else these systems may infer, what those inferences may be used to decide, and whether users will get a meaningful way to challenge them.

Whether age-guessing remains a narrow safeguard or becomes part of a wider pattern of platforms drawing conclusions we did not choose to share is a question still being worked out.

Our AI chats can surface as evidence

Read more

Every month, the Librarian Hotline answers a reader’s question, and those pieces live in the paid library. I’m closing out August by taking the paywall down on one of them, so everyone can enjoy a full edition. If you’ve been wondering whether the Hotline is for you, start here, and pass it on!

The reference interview is a core practice in library science. When a patron approaches the desk and says “I need information about climate change,” a librarian doesn’t pull a stack of books and call it done. They ask questions: What are you trying to understand? What do you already know? What will you do with what you find? The conversation is part of the work, because the quality of what we find depends on the precision with which we’ve named what we’re looking for.

AI breaks this contract. A model trained on enormous amounts of human language is extraordinarily fluent at sounding authoritative regardless of how well we’ve framed our question. It will answer when the question is underspecified, when context is missing, when what we need isn’t quite what we asked for. Unlike a librarian, it rarely tells us we’ve come in with the wrong question. It will just answer the one we gave it, confidently and completely, and leave the gap between what we asked and what we needed entirely invisible.

The breakdown happens in what we do before and after the prompt, not in the prompt itself.

Reader Question

“I’ve been using AI at work for about six months. I ask questions, use what comes back, and move on. But something has started to bother me: sometimes I can’t fully explain decisions I’ve made with AI’s help, or I’ll use an AI-generated summary and find out later it missed something important. A coworker told me I’m just ‘prompting wrong,’ but that doesn’t feel like the whole answer — I follow prompt engineering best practices. What’s going on?”

— Dev P.

Librarian Answer

Your coworker is addressing a symptom. Prompting technique matters, but what you’re describing runs deeper: the cognitive work of evaluation, interrogation, and synthesis got handed off to the model, and it wasn’t equipped to do it. AI can produce fluent, confident-sounding output without understanding your context, your stakes, or what “right” means for your specific situation. That judgment belongs to you, and it doesn’t happen automatically just because you’re using the tool.

When We Outsource the Thinking

There’s a version of AI use that looks like research but functions more like delegation. We bring a question, receive an answer, and move on. The output arrives formatted and fluent, which triggers the same cognitive response we’d have to a well-organized report: it seems finished and authoritative, so we treat it that way.

What disappears in that exchange is cognitive friction, and with it goes the understanding that friction was building. Cognitive friction is the effortful feeling of working through a problem rather than accepting the first answer that arrives. It’s what happens when we read a source and push back on a claim, when we try to explain something and realize we can’t quite do it yet, or when two conflicting accounts force us to decide which one to trust. That discomfort is also how we build the capacity to hold information well, to know when to trust a conclusion and when to keep pulling the thread

A 2011 study published in Science found that when people know they can look something up, they’re less likely to retain it - the brain treats external systems as a kind of storage offload. The effect may be more pronounced with AI, which doesn’t only retrieve information but synthesizes and presents conclusions. We’re not even doing the organizational work of sorting through results anymore; the model is. And when we consistently let it, we risk not developing the judgment that comes from doing that work ourselves.

Dev’s situation reflects this directly. When we can’t explain a decision we made with AI’s help, the reasoning may never have been built in the first place: a conclusion arrived, got used, and the thinking that would let us defend or interrogate that conclusion never took place. What looks like a productivity gain in the short term carries a real risk, and over months and years of consistent practice, the pattern recognition and critical judgment we rely on can erode in ways we don’t notice until we need them.

The gap between “I have an answer” and “I understand this” widens every time we skip the friction, and no amount of better prompting narrows that distance. Output quality matters, but the deeper problem is the thinking we bypass in order to get there faster.

Why Think With AI at All

The case against passive AI use is also the case for something better.

A model trained on the breadth of human knowledge can hold a vast amount of context simultaneously and draw on it in response to what we specifically bring. When we bring something real — a problem we’re stuck on, a position we’re not sure we believe, a decision we can’t seem to make — the model has something to work with, and what comes back is material for thinking rather than a substitute for it.

Used this way, AI can surface a counterargument we hadn’t considered, push back on an assumption we didn’t know we were making, or expose a connection across two bodies of knowledge that might otherwise be difficult to find. A researcher who asks a model to challenge their hypothesis before they finalize their analysis catches problems earlier. A manager who uses AI to pressure-test a proposal before presenting it walks into the room with stronger reasoning. A writer who asks for the weakest point in their argument and then rewrites that section themselves produces better work than either step would produce alone.

In none of those cases did the model do the thinking. It created conditions for better thinking to happen.

This is also what accumulates over time in the opposite direction from passive use. Instead of eroding judgment by outsourcing it, active engagement builds it, because every time we interrogate an output, test it against what we know, and decide what to do with it, we’re doing the cognitive work that makes us sharper at the next decision. Passive use and active engagement compound in opposite directions. The question is which one we’re practicing.

Thinking with AI is less about caution and more about return on investment. When we stay in the room as active participants, the work tends to get better — and so does our capacity to do it.

The Confidence Problem

AI models don’t know when they’re wrong in a way that flags itself clearly. Incorrect information arrives in the same confident register as correct information. The model isn’t hedging based on its actual certainty: it has no epistemic state to draw on. It’s producing the most statistically plausible continuation of our prompt, which is a fundamentally different operation than verifying a fact or reasoning through a problem.

When a doctor says “I’m not sure, let me check,” that uncertainty is meaningful information. It tells us something real about the state of their knowledge. An AI doesn’t have that mechanism. It can be instructed to say “I’m not certain,” but that phrase doesn’t correspond to any internal state of knowing or not knowing. It’s just another output pattern, produced in some cases and not others, without consistent correlation to whether the information is accurate.

The missing summary detail Dev mentions is one of the most common and least visible ways this plays out. When an AI summarizes a document, it’s selecting for what sounds like a coherent summary, which isn’t the same as selecting for what we need to know. A model summarizing a research report tends to foreground the headline findings, the ones that read like conclusions, and compress or drop the methodological detail that would change how we interpret them — not because it made a judgment call, but because that detail doesn’t pattern-match to “key finding.”

The same dynamic appears with contested or time-sensitive questions. Ask an AI about an ongoing legal case, a company’s current leadership, or a scientific question still being actively debated, and it will often produce an answer that sounds settled, because the model is trained to produce answers, and answers sound confident by nature. The confidence is structural, not earned.

Healthy skepticism toward AI output isn’t excessive caution. It’s an appropriate response to a tool that can be wrong without any indication of error.

The Reference Interview, Applied

A practiced reference librarian doesn’t answer the question as asked. They answer the question behind the question — the one the patron actually came in with, which is often different from the one they put into words.

The reference interview works because the librarian and patron are thinking together toward something neither could reach alone. Asking AI for an answer is a transaction: we know what we want, we’re just retrieving it. A thinking partnership is something different — we arrive with something unresolved and use the conversation to work through it, with the model as a collaborator in the process rather than a dispenser of conclusions.

What the shift looks like in practice:

• Asking AI: “What are the main causes of employee burnout?”

• Thinking with AI: “Three people on my team seem fine, two seem checked out, and the only thing that changed two months ago was switching to asynchronous communication. I keep assuming it’s the structure, but what else should I be considering before I draw that conclusion?”

• Asking AI: “Summarize this article.”

• Thinking with AI: “I want to reference this piece in something I’m writing, but I realized I can’t articulate its actual argument — I just came away with a general vibe. Can you help me distinguish what the author is claiming from what I’m inferring?”

• Asking AI: “Should I take this job offer?”

• Thinking with AI: “I’ve made a pro/con list for a new job and it didn’t help. I think the problem is that I’m weighing things I can measure — salary, title, commute — but the hesitation is about something I can’t put my finger on. Can you help me figure out what’s actually at stake for me here?”

• Asking AI: “What’s the difference between a Roth IRA and a traditional IRA?”

• Thinking with AI: “I understand the tax difference between a Roth IRA and traditional IRA in theory, but every time I try to apply it to my actual life I freeze because I don’t know what my future tax bracket will be. Can you help me figure out what I’d need to know about my own situation to make this less abstract?”

• Asking AI: “Is this a good business idea?”

• Thinking with AI: “I’ve been sitting with this idea for months and everyone I’ve told has been encouraging, which somehow makes me more uncertain. I can’t tell if I believe in it or if I’m just attached to it. What’s the actual difference between those two things?”

• Asking AI: “What should I say in this performance review?”

• Thinking with AI: “I’m writing feedback for someone who has real skill but a pattern that’s hurting the team. I keep cushioning it and I can’t tell if I’m being diplomatic or dishonest. What’s the difference, and how do I stay on the right side of it?”

• Asking AI: “How do I know if I should change careers?”

• Thinking with AI: “I’m good at my current job and drawn to something else, and the fact that both things are true is what makes this hard. I don’t need help identifying my options — I need help figuring out what kind of decision this even is. Is this a risk calculation or a values question?”

• Asking AI: “What’s a good way to handle this conflict with a coworker?”

• Thinking with AI: “Something’s off with how credit is landing on a project I’m sharing with a colleague, but I keep going back and forth between ‘this is a real problem’ and ‘I’m being oversensitive.’ Can you help me think about what evidence would separate those two?”

• Asking AI: “Write a follow-up email after a networking meeting.”

• Thinking with AI: “I’m writing a follow-up to someone I’d like to stay in touch with, and every version I draft sounds like a version of me that isn’t quite right — either too eager or too casual. Can you help me figure out what I actually want from this connection so I can write something that sounds like me?”

The pattern is the same in every case. Asking AI hands the model an open question and accepts whatever frame it chooses. Thinking with AI brings the actual situation, names what we’re trying to figure out, and asks the model to work inside our reasoning rather than substitute for it. What comes back is still just output, but it’s output we’re in a better position to evaluate.

What AI Cannot Supply

Everything an AI model knows about our situation is what we’ve told it in the conversation. It has no access to the history of a working relationship, the political dynamics of an organization, the constraint that makes an otherwise sensible solution unworkable, or the value judgment that has to land at the end of an analysis — unless we put those things on the page. That’s not a design flaw. It’s just the nature of the exchange, and it means the quality of what we get back is inseparable from the quality of what we bring.

This is why the thinking can’t be fully delegated. The model generates plausible text. But we determine whether that text is true, complete, and appropriate for the actual decision we’re facing. Those are different jobs, and only one of them requires knowing what’s really at stake.

That practice, repeated over time, is also what protects something harder to recover once it’s gone. The judgment we build through sustained cognitive engagement develops through use and atrophies when we stop. It degrades gradually, in the same way any skill erodes when a tool starts doing the work instead.

The people who get the most out of AI over the long run aren’t the ones who use it the most. They’re the ones who stay in the room while they use it, who treat every output as something to be interrogated rather than accepted, and who bring their own critical thinking to the exchange. AI can be one of the most powerful cognitive tools we’ve ever had access to. Whether it functions that way depends entirely on what we bring to it.

Quick summary:

An AI company has been buying up used books by the million and destroying them, scanning the pages for training data and pulping what’s left. Court filings unsealed this year describe the program and an internal note asking that the work be kept from becoming known. A court has ruled that buying and scanning the books this way is legal, and once the work is done, nothing survives to show a book was ever there to lose.

Key takeaways:

→ The books are bought through anonymous middlemen, so the sellers filling the orders rarely know where their stock is headed. From every angle the buying looks like ordinary commerce, which is exactly what keeps the destruction from being seen.

→ Book destruction has a long history, and it changes each time the technology of copying changes. It has shifted twice before. This is the third shift, and it looks nothing like the book burnings most of us picture.

→ This shift is set apart by leaving nothing behind. The words are kept, the object is discarded, and no record says which titles were taken, so the loss can be neither proven nor traced to anyone who might answer for the harm.

→ That reaches past books to anyone who wants to know things firsthand. When the only surviving copy is sealed inside a system no outsider can consult, verifying what it says becomes impossible, and we are left trusting whatever summary we are given.

“We don’t want it to be known that we are working on this.” The sentence appears in an internal Anthropic planning document made public through court filings in Bartz v. Anthropic, the copyright class action that three authors filed in 2024. Anthropic, the maker of Claude, called the program Project Panama. Its stated mission fit in one sentence: “Project Panama is our effort to destructively scan all the books in the world.”

Destructive scanning means cutting a book from its binding, feeding the loose pages through a scanner, and discarding the book afterward. According to the court, Anthropic spent millions of dollars buying millions of print books, often used. Its service providers removed the bindings, cut the pages, scanned them into searchable digital files, and discarded the books. Anthropic kept the scans in an internal digital library and used books from that library to train the AI systems behind Claude.

The lawsuit initially concerned another source of Anthropic’s library: more than seven million pirated book files the company had downloaded. The court treated the two acquisition paths differently. It held that Anthropic could lawfully buy print books, convert them into searchable digital files for internal use, and discard the physical copies; because the resulting files remained inside the company rather than being redistributed, that practice was fair use. It reached the opposite conclusion about the books Anthropic had downloaded from pirate sites and retained.

Companies announce the programs they are proud of; Anthropic tried to keep this one invisible. The January unsealing supplied the project’s codename and the instruction to stay silent. The instruction anticipated what the company did not want authors, booksellers, and the reading public to see: books bought by the million, cut apart, scanned, and discarded. Anthropic didn’t publicly announce Project Panama before court records made it public.

The silence has no precedent. Books have been destroyed for as long as they’ve been made: by conquering armies and offended churches, by censors with lists and mobs with torches, by floods and fires, and by budgets that let the roof leak. Some of it was fast and public. More of it was slow and official. All of it, loud or slow, could be recognized for what it was. A person watching knew that books were being lost. History kept what record it could. But what’s happening now carries no such signature. From the outside, nobody could tell that books were being destroyed at all.

Inside Project Panama

Anthropic hired a man named Tom Turvey in February 2024 and gave him a mission the court record repeats in one sweeping phrase: obtaining all the books in the world. Turvey came to the job from Google Books, the project that spent the 2000s digitizing library collections on machines engineered to turn pages gently, so that every book survived its own scanning. Anthropic took the opposite approach. Gentle machinery never entered the plan. Over roughly a year, the company spent tens of millions of dollars buying millions of print books, sheared off their spines, fed the loose pages through high-speed scanners, and pulped what remained. Vendor proposals in the court records described converting up to two million books in six months, some eleven thousand a day.

Destroying the books solved a financial problem first. A bound book must be scanned page by page, slowly and at cost. Once cut apart, the same volume becomes a stack of loose paper that can run through a sheet feeder at speed. Across millions of volumes, the difference in output determined the method.

The legal significance emerged later. In June 2025, Judge William Alsup ruled on the authors’ claims. For the books Anthropic had bought and destroyed, he found the copying to be fair use: the rule in American copyright law that allows limited copying without an author’s permission (the way a critic can quote a novel in a review). His reasoning turned on replacement. Anthropic bought one physical copy, made one digital copy, and destroyed the original—so the number of copies in the world never grew. In the law’s eyes, the scan simply took the book’s place, a change of format. If Anthropic had kept both the book and the scan, the number of copies would have doubled and the fair-use argument would have weakened. Pulping the originals helped make the copying legal.

The pirated downloads fared differently in the same ruling. For those more than seven million files, Alsup rejected Anthropic’s fair-use defense. He emphasized that Anthropic had built a permanent, general-purpose library it expected to keep indefinitely, not a temporary collection for a defined training task. “None is even offered here except for Anthropic’s pocketbook and convenience,” he wrote.

The ruling left the company facing a trial over damages, and Anthropic settled instead: the company agreed to pay $1.5 billion. A judge granted the deal final approval on July 20, the largest copyright class settlement in American history. Under the deal, Anthropic will delete the pirated digital files. Critically, the deal concerns those files, not the millions of physical books the company bought and pulped. Copyright protects the text of a book—the creative expression of an idea—not the paper on which those words were printed. Once Anthropic lawfully owned a physical copy, copyright law generally didn’t prevent it from destroying that object. The books could be destroyed without creating the kind of copyright violation at issue in the case.

But books don’t enter a library by themselves: somebody had to sell the company all those books. The court records describe Anthropic purchasing through vendors. This spring and summer, booksellers across Europe described the other side of such a trade: unusually large, eclectic orders placed through intermediaries, with the ultimate buyer unnamed. Tomás Kenny of Kennys of Galway called one order for books “bananas”—a mix of titles no library would plausibly assemble. Whether any particular order came from Anthropic cannot be established from outside the transaction. Nor is Anthropic the only possible customer: other AI companies have also been reported to be acquiring books at industrial scale.

The reported market is built to keep the buyer at a distance. Brokers can aggregate inventory and manage bulk orders without disclosing who is ultimately acquiring the books. ISBNdb, a book-data company, briefly advertised a prospective book-sourcing service for AI developers that promised confidentiality; its marketing explained the appeal bluntly: “‘AI company destroys two million books’ is not a headline that generates sympathy.” (After reporting drew attention to the page, ISBNdb removed it and said the proposed service had never been launched.) The result was not merely secrecy about the buyer, but uncertainty about the fate of the books.

Annihilation, then spectacle

What Anthropic is doing belongs to a history far older than the company. Rebecca Knuth, a professor of library science, named the practice in 2003: libricide, the systematic destruction of books and libraries, usually carried out or authorized by a government. Her subject was the twentieth century’s state-sponsored campaigns, but the practice runs back as far as writing does. The scenes that come to mind of this are the same few: students feeding bonfires in Berlin in 1933, Sarajevo’s national library burning under siege in 1992. But behind those scenes, the record is wider and stranger than they suggest. Bonfires were rarer than the memory of them. Most destruction arrived slowly and with permission, through purges, censors, wars, and simple neglect.

A strict reading of that definition would leave Anthropic out, reserving libricide for the destruction of entire libraries. But that distinction collapses here. The world’s secondhand book trade functions as one enormous collection, scattered across thousands of shops and sellers with no central address. Buying it up by the pallet and pulping it empties a library all the same, just one whose shelves span continents.

Destroying a book has meant different things in different centuries. The difference has always come down to copying, because how books are copied decides how many of any one book exist. When copies are scarce, destruction can erase a work from existence. Once copies are everywhere, destruction can only send a message about one. To me, that line sorts the history of libricide into eras, and what distinguishes each one is what its destruction leaves behind. The result is my own framework, a chronology by residue that I haven’t found anywhere in the scholarship: two eras completed, and a third that has just begun.

For thousands of years, every copy of every book was made by hand. A single volume could take a scribe months, so most works existed in a few manuscripts, and some in only one. Under those conditions, destroying the object destroyed the work, completely and forever. I call this the first era: the era of destruction as annihilation. The word descends from the Latin ad nihil, meaning to reduce or bring something to nothing. In this era the meaning was literal. When Diego de Landa, a Spanish friar in colonial Yucatán, burned twenty-seven Maya codices in 1562, the texts inside them went to ash. No copies of them existed anywhere on earth, so entire bodies of Maya history and belief ended in one afternoon’s fire. The library of Alexandria met the slower version of the same fate, declining through centuries of purges and neglect until its losses ran past counting. What the first era left behind was dust, and one thing more: knowledge of the loss. Contemporaries recorded what had burned. We can still mourn the codices, because the one thing annihilation couldn’t destroy was the memory that the books had existed.

The printing press ended that era within a century of its invention. Once a title could exist in hundreds or thousands of identical copies spread across cities and countries, fire lost its reach. Burning a book now destroyed only an object, since the work lived on in every other copy, safely out of range. Destruction continued anyway, though, serving a different purpose entirely. I call this the second era, the era of destruction as spectacle (a word descended from the Latin spectare, to watch). By the twentieth century, watching had become the entire point of burning a book. The clearest case is the Nazi bonfires of May 1933, when German students burned tens of thousands of volumes in public squares, in front of rolling newsreel cameras. Almost none of the works truly died in those fires, because the titles on the pyres existed in editions across Europe and America (many of which remain in print today). Erasing the books was never the goal, since erasure had stopped being possible. The fires existed to be seen, a threat performed first for the crowd in the square and then for everyone who watched the footage. What the second era left behind was the opposite of ash: photographs and visuals of fires that consumed real books but reached nothing beyond them. Once a book existed in enough copies, burning one no longer removed it from the world. It announced that the book had no place in the world to come.

Destruction as disappearance

Measured against those two eras, what Anthropic is doing fits neither. In the first era, destroying the object meant losing the work. Anthropic’s scanners preserve every word, so the works survive. In the second era, the objects were beside the point and the burning was public theater. This time the objects are destroyed by the millions, and the destruction says nothing at all; it’s not a message but a method. The company ordered the operation kept out of sight. A destruction that erases no text and performs for no one, run at industrial speed, matches neither pattern. The technology of copying has crossed another threshold, the way it did when the press replaced the scribe. What it means to destroy a book now has changed again to match. We’ve entered the third era.

What exactly went into the scanners is the question nobody outside can answer. The unsealed documents show the program favored what its leader called “less common” books, harder-to-find titles over mass-market ones, without ever defining where less common ends. After the claims went viral this summer, the fact-checking site Snopes investigated whether rare books were being pulped. Anthropic told Snopes that “none of our data acquisition programs buy and destroy ‘rare’ or ‘antiquarian’ books.” The assurance can’t be tested from outside, because no list of what was bought has ever been made public.

Anthropic’s own planners estimated the world has about 130 million distinct books. The program destroyed millions of copies, most of them ordinary used books with plenty of surviving duplicates. Even where duplicates survive, a scan keeps only the words. A physical copy carries evidence too, like a censored paragraph that marks one printing apart from another, or an owner’s name inked inside the cover. But the deeper danger sits in the margin nobody can see. Anthropic aimed for “uncommon” titles while recording nothing public about which copies were destroyed. For a book surviving in only a handful of copies anywhere, one bulk order can potentially take the last one. Whether that has already happened is a question no one on the outside knows for sure. The impossibility of answering what was destroyed is what’s new.

The buying also leans toward older books, for a reason that has nothing to do with rarity. Since around 2022, text written by AI has spread across the internet, mixed in with everything people write and mostly impossible to tell apart. That creates a problem for the companies training new models. Feeding a model text written by other models tends to degrade the results, so the builders want sources guaranteed to be human. A book printed before the technology existed carries that guarantee on its copyright page. Old print has become a raw material, valued for the one quality the internet can no longer promise.

One more feature separates this era from the last one. Spectacle-era destruction wanted an audience; this destruction wants the opposite. The buyer is after the text and has no message to send. In fact, attention to its process can only slow the buying down. Taken together, the features of this era line up: the works survive inside a black box, the objects vanish by the millions, no public record exists, and the operation itself prefers to work in secrecy. I call this the third era, the era of destruction as disappearance. The word rests on the Latin apparere, to come into view, with a prefix that reverses the motion. A disappearance is a departure from visibility. The word fits this era at every layer, from the unseen sales to a program that only appeared when a court forced it into view.

Every one of these books, preserved to the letter, can now be read by no one. The scanned texts sit in Anthropic’s private collection, a library with no reading room and no public catalog. Models trained on that collection are built to avoid quoting it at length, because reproducing long passages for users is the copyright violation no court has excused. So the machines answer questions about books without ever showing the books themselves. When someone asks a model about a title that survives only in that collection, what comes back is a summary, written in the model’s words, with the original nowhere in reach.

That arrangement changes something basic about how we can know things. Checking a claim against its source is the foundation of thinking for ourselves. When we can pull the book, find the page, and read the passage in context, we get to judge whether the summary was faithful and whether the quote meant what someone claims it meant. Remove the book and that judgment has nowhere to stand. We’re left taking the summary on trust, with no way to confirm and no standing to doubt. Whatever the model says the book said becomes, for all practical purposes, what the book said. That’s a transfer of authority from the page to the tool, from a source anyone could check to an answer nobody can. The transfer happens one unreachable book at a time.

Maddeningly, every piece of this design is legal. Each one also removes a question that used to have an answer. The anonymous purchasing means nobody can list which books were destroyed, so nobody can rule out that some were the last copies anywhere. With the collection sealed, nobody can compare a stored text against the printed original it replaced, so even perfect fidelity can never be shown. As for whether the buying ever stopped, nobody outside knows that either, since the only reason anyone knows it started is that a piracy lawsuit dragged the records into the open. None of this took a conspiracy, only ordinary business decisions about what to disclose, all of them legal and every one of them closing a door.

One more thing disappears along with these books: the ability to mourn them. The previous era’s destruction left survivors who could testify to what was lost. Alexandria’s losses were lamented for centuries. Sarajevo’s librarians catalogued what their fire took. The third era leaves no one who can even compile the list. A loss we can name is a wound. One we can't name is just a world grown slightly smaller, with nothing to point to and no way to prove it. A librarian would describe the situation in the profession’s terms: no accession record and no deaccession record, a transaction that leaves no ledger for anyone to audit. Put simply, they’ve created a system in which they can’t be held accountable for the system they’ve created. That understanding defines destruction as disappearance: the loss was built, from the start, to be impossible to establish.

What can still be known

The concealment had one structural weakness: a program that buys millions of books needs sellers, and a company that breaks the law can be sued. Sellers and courts are the two doors into the secrecy — sellers see every order even when they can’t see the buyer, and courts can compel what the company won’t volunteer. Everything now known about the program came through one door or the other. Authors sued and forced the internal records open. Booksellers noticed matching orders across two countries and brought them to reporters; one worked with the journalists at 404 Media to hide a tracking device inside a shipment of books, and the tracker led to an Amazon warehouse outside Las Vegas — evidence that a retail giant appears to be running a scanning program of its own. Within weeks, at least one supplier that had been advertising bulk books for AI training pulled the offer. No regulator opened either door. Every disclosure came from the people the company bought from or the people it answered to in court.

Some sellers went further than noticing. Once Tomás Kenny, whose Galway shop had received one of the strange 5,000-book orders, worked out where books like his might be headed, he said publicly that Kennys wanted no part of scanning aimed at extracting intellectual property. His shop had been the second in the world to put its books online, back in 1994; it now became one of the first to refuse the trade that takes books offline for good. Kenny could refuse because he had worked out the destination, which is exactly the discovery the brokers’ anonymity exists to prevent. A trade that hides its purpose from its own suppliers has already answered the question of whether the suppliers would approve.

The same kind of attention is available to the rest of us, because most of us eventually stand over a box of books deciding where they go, after a move or the clearing of a family house. That box is where all of this arrives at our own doorstep. A seller, or any of us selling to one, can now ask a question that has a good reason to be asked: where do these books go next? The anonymity that keeps this market running survives only as long as nobody asks. For material that might be scarce, like a town history or a box of local records, the open market is no longer the safe default. A library’s special collections desk exists for exactly that kind of donation. And the same habit applies at the other end of the pipeline: when a model summarizes a book for us, we can treat the answer as a starting point and go find the book itself, while findable copies still exist. Each of those is a small act of keeping track.

That kind of record-keeping has been the difference between the eras all along. Each era’s destruction left a residue, a testament of a kind. The first era of annihilation left ashes and knowledge of the devastation. The second era of spectacle left photographs of the fires and the threat those fires were meant to carry. Ours, this third era of disappearance, leaves no ash, no image, and no list. But what’s still open is the record itself: whether one gets kept, and whether anyone outside can access it at all.

The three eras are a framework I built for this piece, and I want to take it further (it’s such fascinating stuff!): a full treatment of how book destruction has changed across five thousand years and what the third era asks of the people living in it. Before I build it though, I’d love to get a temperature read on the format you’re most interested in.

The poll's anonymous, so if you'd want this for a group or an organization, leave a comment, reply to this email, or send me a DM on Substack so we can connect.

You might also like:

• The Libraries That Weren’t Supposed to Exist: Basement libraries and buried archives: a brief history of reading against the state.

• Who Benefits From Us Not Knowing the Full Picture?: How information asymmetry shapes what we buy and believe.

• How One Rescue Mission Turned into the Global Record of Human Heritage: Inside the preservation system that classifies, catalogs, and protects humanity’s cultural and natural inheritance.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides.

Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

That experience can produce a private sense of failure. We may conclude that we read too slowly, missed essential context, or lacked the discipline to keep up. The problem often lies elsewhere. Public information now arrives through systems that reward speed, repetition, emotional reaction, and constant return. When the volume exceeds the time available for verification and reflection, the difficulty is by design.

Information literacy begins there. It doesn’t ask us to consume every available piece of information out there, but gives us ways to assess origin, context, evidence, timing, and purpose. Those habits become especially valuable when a public event is surrounded by so much material that no single person can follow the whole field.

In early 2018, the writer Michael Lewis watched the State of the Union address with Steve Bannon, who had left the Trump White House the previous summer. Bannon described his approach to dealing with the media in an unusually direct phrase: “The Democrats don’t matter. The real opposition is the media. And the way to deal with them is to flood the zone with shit.” The phrase became shorthand for a familiar political and media tactic: saturating the public sphere with claims, controversies, provocations, and announcements so that sustained attention to any one damaging development becomes harder to maintain.

A flood of information doesn’t depend on every item being false. A fast stream can contain accurate reporting, misleading fragments, official statements, manufactured drama, and sincere public response. But the volume changes how all of them are received. Once we have more material than we can assess, the order of arrival begins to govern attention more than evidence does.

The language of flooding also explains why ordinary habits often fail during such periods. Opening one more tab doesn’t solve the problem, nor does saving every link we come across. And sharing the most alarming claim before its context is established can deepen the same conditions that made it hard to think clearly in the first place.

A practical response lies in preparation and boundaries. A flood doesn’t disappear simply because we understand that a flood is happening. We can, however, decide how much enters our day, when we examine claims, what we preserve, and which sources and institutions we trust to help watch the wider field. Those boundaries preserve our ability to follow public events without letting volume choose the order of our attention.

How the flood works

The phrase “flood the zone” comes from American football. One team tries to move the ball by throwing it to a teammate. Sometimes, it sends several teammates into the same small part of the field, each ready to receive the throw. The other team can’t watch everyone at once, so one teammate may have a clear chance to catch the ball.

Information flooding puts similar pressure on attention. Journalists and readers have limited time to investigate what they encounter. When many claims, controversies, announcements, and reactions arrive at once, they can’t give each one equal scrutiny. A consequential story may then be crowded out, while a less important or more provocative item captures attention. A stream of loosely related claims can also consume the hours that reporters and readers might otherwise spend examining a development with lasting consequences.

This tactic isn’t new. Governments and organizations have long been accused of releasing unfavorable news when public and press attention is already divided. Crowded news cycles can make sustained scrutiny harder to secure. Digital platforms intensify those pressures by distributing material continuously, while systems built to maximize engagement can give emotionally charged posts greater visibility. Fresh reactions can then seem more urgent than the evidence warrants.

The result can be disorientation. People may lose confidence in their ability to reconstruct what happened and decide which accounts deserve trust. By the time a correction arrives, attention may have shifted elsewhere. A well-reported investigation then competes with a fast-moving argument that generates repeated commentary and intense reaction.

The burden of verification also falls unevenly. Making or reposting a claim often takes less work than tracing the claim to its original source and examining the evidence in context. Comparing independently reported accounts only adds further effort. That difference gives a high-volume publisher an advantage over anyone trying to evaluate each item responsibly.

Generative AI has made one part of that imbalance more pronounced. It makes plausible-looking text, images, audio, and video easier to produce at scale. A coordinated campaign still needs to get material in front of people and make it seem worth trusting. But the supply of material is now easier to expand, and more of it can circulate through the same channels where reliable reporting competes for attention.

Ordinary acts of sharing can also move the flood forward. A screenshot can pass from one group chat to another after being separated from its original source. A video clip can circulate after its beginning or ending has been removed. With each act of sharing, another unexamined item can enter someone else’s field of attention before its meaning has been established.

Recognizing a flooded zone

A flooded zone rarely announces itself. Any one item may look ordinary on its own. The pattern becomes visible when we notice what the accumulation is doing to our attention: which subjects dominate and which questions go unanswered.

Several conditions often appear together:

• One subject suddenly dominates alerts, feeds, conversations, and commentary across many platforms

• New reports arrive faster than earlier ones can be checked or placed in context

• A major story slips from view before its central questions have been answered.

• Corrections may receive less attention than the claims that prompted them

• A sudden controversy coincides with a consequential vote, ruling, executive action, corporate decision, or policy change

A fast-moving controversy may dominate coverage at the same time a government or organization announces a decision that changes public policy or institutional practice. And when that happens, we can examine whether the controversy displaced attention from the decision without assuming that the timing was coordinated.

Some developments provide advance warning. Periods of political consequence (especially elections and major policy developments) often produce concentrated waves of coverage. Keeping a calendar of such events can help us anticipate periods when public attention may face greater pressure. The calendar can also help us decide which sources or developments deserve closer attention beforehand.

A private record can deepen that awareness. We might note a major controversy, the date it appeared, and other public actions occurring around it. Returning later can reveal which event dominated attention and what subsequent reporting clarified. The record turns vague overwhelm into an account of how public attention moved.

Information literacy asks whether a claim is supported by evidence. In a flooded zone, we also examine how the claim reached public attention and which event lost visibility after it appeared. Placing the claim within that wider information environment shows how volume affected what the public had an opportunity to examine.

How to stop a flood

When water rises, sandbags are placed between the flood and a home. They redirect the current before it enters the house, leaving the source unchanged. An information flood calls for a similar response. Trying to absorb every claim and reaction allows the flood to occupy more of the day. But we can place boundaries between the surge and the attention we need for judgment and daily life.

I call those barriers sanity sandbags. These deliberate actions limit how much of a flooded information environment enters our daily lives. They create boundaries between the flood and the parts of our life that need protection (our capacity for judgment and our nervous system regulation chief among them). We can remain aware of what’s happening without giving a relentless stream of information authority over every available moment.

Sanity sandbags can be set before predictable periods of high volume or added after the flood has begun. Each one creates more room to assess what is happening. The added distance helps us decide how to respond without allowing urgency to decide on our behalf.

Sandbag #1: Step away from the current

Create distance from the live feed.

We close the feed and step away until the next planned check-in. Channels carrying emergency instructions or information tied to an immediate responsibility remain open. Distance comes first here, because every later choice depends on having enough attention left to choose deliberately. Breaking news can feel hardest to leave because uncertainty creates pressure to keep checking. But during those periods, reporting is still being assembled. Early accounts may conflict before basic facts have been confirmed. Corrections may not reach the people who saw the first claim. When we return, the reporting may include evidence and context that were absent from the first accounts.

Stepping away during a public crisis can feel like neglecting a duty. Continued exposure can weaken the attention we need to evaluate later reporting. We step back so that we can return with enough attention to compare accounts and decide what deserves further examination.

Sandbag #2: Let the archives do the archiving

Preserve what belongs to your own work.

A flood can make every passing item feel irreplaceable. Once we step away from the feed, another anxiety appears: what if the article or post disappears before we save it? That fear keeps us tied to the current. Each new item can seem to demand a private copy.

The concern reflects a documented problem. A Pew Research Center analysis found that one quarter of webpages available at some point between 2013 and 2023 could no longer be reached in October 2023. Link rot describes one form of digital loss, when a link no longer leads to the page it once did. Some pages disappear entirely. Others change through revision or become difficult to locate as newer material accumulates.

Saving material can keep this public record within reach. The difficulty begins when the urge extends to everything moving through the feed. A personal archive can grow faster than we can organize or revisit the collection. The work of saving begins to resemble the flood itself, with each new article creating another demand on our attention.

Preservation becomes manageable when the labor is divided. We keep the pages tied directly to our responsibilities. Sources supporting our writing belong in a personal archive, and records needed for an ongoing responsibility may also deserve a personal copy. Public pages, however, can sometimes move into institutional archives. The Internet Archive’s Wayback Machine has collected copies of webpages since 1996. Its Save Page Now feature lets anyone request an archived copy of a public page. The End of Term Web Archive preserves federal websites during presidential transitions, when official information can change or disappear. So a page earns a place in our personal archive when losing it would interfere with work already in our care or remove evidence we need to retain. Material of wider public consequence can go to a shared archive, leaving our personal archive small enough to use rather than turning it into another place where the flood continues.

Sandbag #3: Schedule the intake

Choose when the news enters the day.

Once we’ve decided what belongs in our archive, we can decide when the news enters the day. We choose a fixed reading window and keep the feed closed in between. The boundary replaces ambient monitoring with a deliberate encounter that has a beginning and an endpoint.

A live feed presents events in the order they surge into view. A trusted human-edited roundup can place developments beside reporting that explains their context. Reading within a defined window places accounts beside one another, making comparison possible before a new alert redirects our attention.

A fixed reading window also limits repeated exposure. RAND’s review of psychological research notes that repetition can make a claim seem more credible. Encountering the same assertion fewer times leaves less room for familiarity to influence our judgment before we’ve examined the evidence.

Sandbag #4: Let the peak pass

Give the story time to become clear.

A breaking report can create pressure to decide immediately what happened and how we should interpret the event. But the pressure often arrives before reporting can answer either question. The conclusion we form under that pressure may rest on evidence that remains incomplete.

For developments that don’t demand immediate action, consider the first reporting as provisional. Avoid building a firm view around material created during the first rush of coverage. The first version of a story can contain a partial account without showing the sequence that gave an event its meaning. Later reporting can place that account within a fuller record, showing which conclusions the evidence supports and which questions remain open.

The pause protects memory as well as judgment. Psychologists use the term continued influence effect for the way misinformation can continue to affect reasoning after a correction. An early claim may influence how we understand an event after later reporting has shown the claim to be incomplete or false. The first account becomes a reference point for the information that follows. Waiting before we commit to an explanation gives more complete reporting a chance to arrive before that account settles into memory.

Some floods consist of accurate developments arriving too quickly for close attention. A series of executive orders or court decisions can overwhelm attention even when the reporting is sound. Waiting doesn’t mean ignoring events that affect our lives or delaying action when time matters. But the pause gives us room to decide which developments need a response now and which stories will become clearer as reporting develops.

Sandbag #5: Stop carrying the water

Pause before passing a claim along.

During a surge, passing something along can feel like a public duty. We may want to warn someone about a claim, or ask for help making sense of what we’ve seen. Before sharing, though, we can ask whether the material has a source the recipient can inspect for themselves. We can also ask whether we have enough context to explain what the material shows and what remains unresolved. When either answer is uncertain, leaving the material unshared keeps an unverified claim from entering another person’s day through someone they trust.

Pausing preserves the difference between repetition and corroboration. A claim can appear to have support when several posts repeat the same account without adding independent evidence. Sending the original source allows the recipient to examine the evidence directly. Naming any unresolved questions keeps our message from sounding more conclusive than the reporting supports.

The pause also changes our own relationship to the flood. We spend less time attached to every fast-moving item. We also spend less time following the reactions that appear after we share a message. People close to us receive fewer questionable claims through our messages. Each decision not to forward closes one route through which the flood would otherwise expand.

Sandbag #6:

Share the work of paying attention.

A flood becomes harder to manage when one person tries to watch every place where water may break through. When several areas need attention, people can divide the watch and alert one another when conditions change. Public information presents a similar problem. No one person can follow every court decision or international development with equal care. Dividing the watch lets each of us follow a smaller part of public life more closely.

We can divide that work among people whose judgment we trust. One friend may follow decisions from the courts. Another may understand events in a region through family ties or knowledge of the local language. Each person can share reporting that identifies where the information came from. They can also explain what the reporting has established and what remains uncertain.

A shared news watch can begin with one other person who follows a different area. Each person sends reporting when a development affects an issue the other has chosen to follow. A larger group can agree on a limited set of subjects and use one shared channel for reports. An incomplete watch can still extend what each member can follow. The arrangement gives us access to knowledge beyond our own attention without recreating the continuous feed inside the group.

What we take back

Every instance of the flood assumes the same thing about us: information is something done to us, arriving at a pace someone else sets and in an order someone else arranges. The assumption runs so deep that people announce the strategy openly, the way Bannon did, without fear it will stop working once we know. The people who flood the zone are attacking our attention, and deciding for ourselves how we take in the news is how we refuse them. Each time the water rises, we choose whether to be the passive audience the entire strategy depends on. The zone will flood again, around the next election or the next manufactured uproar, but putting the sanity sandbags in place can begin as early as tonight.

Every flood is built on the same wager: that the people underneath it will stay scattered, exhausted, and alone. Meeting the next one on our own terms, at our own pace, in the company of people we trust, is how they lose that bet. The water will keep rising, but it no longer gets to decide how we move through it.

You might also like:

• A Field Guide to Information Overload: How to navigate an environment built to distract us.

• The Unspoken Rules of Information Etiquette: Six practices for how to treat each other when we exchange information.

• The information diet isn’t enough. What we need is an information village.: What changes when we go beyond optimizing our consumption and start building shared infrastructure.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides.

Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

This week:

• A new European law requires companies to say when people are dealing with AI, whether a chatbot on a service line or a synthetic image in an ad.

• California is weighing limits on AI chatbots used in place of therapy, as more people bring their mental health to a machine.

• As a typhoon approached China, AI models built on decades of weather records forecast its path alongside the traditional systems.

Europe starts labeling its AI

What happened

As of August 2, a new European law gives people the right to be told when they are dealing with artificial intelligence. If a chatbot answers a complaint, it now has to say that no human is on the other end. An image or video in an ad that was generated or altered by AI has to carry a label marking it as synthetic. The law doesn’t ban AI from ordinary commercial life; it requires disclosure, so people can know when a machine is involved and decide how much to rely on it.

The rule covers more than the well-known AI companies. A call center that uses AI to read a caller’s mood has to disclose that at the start of the call. Even commercial uses of AI, including scheduling appointments, handling correspondence, or negotiating contracts, will have to be declared. Companies that break the rule can be fined up to 15 million euros, about 17 million dollars, or three percent of their total global revenue for the year, whichever is larger.

The rules are live, but enforcement is unlikely to be immediate or uniform. The EU’s AI Office took longer than expected to produce key compliance guidance, and the bloc’s 27 member countries are at different stages of setting up the systems meant to supervise the law. By December, AI model providers have a transition period to label synthetic audio, images, video, or text in a machine-readable form, so other software can recognize it as machine-made. Regulators acknowledge that this is technically difficult to get right.

Why it matters

This is the first time a major government has required AI to identify itself across everyday commercial life. The reasoning rests on a gap: a person who assumes a human wrote the reply or took the photo has no reason to question what they see. A label closes that gap in a usable way. Someone told that a chatbot is handling their complaint can decide whether to continue, look for another channel, or ask for a human instead. A shopper who learns a product photo is synthetic knows the image may prove little about what arrives in the box.

Europe has changed the internet from inside its own borders before. When its data-protection law, the GDPR, took effect in 2018, companies worldwide rebuilt how they handled personal data rather than run a separate European version of every product. Cookie-consent pop-ups became a familiar part of the web, reaching far beyond Europe. The AI labels may spread the same way, as companies decide it is simpler to build one disclosure system than separate European and non-European versions of their products.

Two things could still keep the law from working. If a label appears on nearly everything, from a spell-checked email to a lightly filtered photo, people may stop seeing the labels at all. One industry group has already warned of exactly that, calling it banner blindness—the same fate that overtook the cookie pop-up. The other risk is that some companies would rather scale back their AI, or bury it deeper, than advertise how much of their work a machine now does. Either way, the law could stay on the books while losing the effect it was meant to have.

What it means for us

The first change many of us notice may be small but concrete. A voice on a service line may say it is automated before we start, and an ad may note that its flawless image was built by AI. That knowledge is something we can act on: we can ask the automated line to pass us to a person, or treat a too-perfect product photo with more caution once a label reveals it was built by a machine.

The labels may reach us wherever we live, since companies adding them may choose to build one global version of a product rather than a separate one for Europe. For each of us, the disclosures add one more way to know who, or what, we are dealing with before deciding how much to rely on an answer. A missing label proves little for now: enforcement is uneven, some uses fall outside the rules, and companies outside Europe may not use the same disclosures.

Bottom Line: The label could also start working in reverse. Once disclosure becomes routine, the absence of AI could turn into something a company can sell, much as food labeling helped turn a word like organic into a claim shoppers will pay extra for. If that pattern repeats, ads may soon announce that a support line or a photograph involved no AI at all, with a higher price attached to the promise. A law written to make AI visible could end up making human work the selling point.

California weighs a limit on AI therapy

Read more

Finding those materials involves more than searching for the right words because different kinds of information are organized and made discoverable in different ways. A library catalog contains structured records for books, journals, media, digital resources, and other materials, and it may also link to or include records for archival collections. An archive may organize thousands of records according to the person or institution that created or accumulated them, then describe groups of records rather than every individual document. A government data portal organizes measurements according to categories such as geography or reporting period, while a scholarly database helps us find research distributed across many publications. Each system is designed around different kinds of information and discovery, which means we may need to search each one differently.

That difference becomes easy to miss when many of these systems can be reached through the same general interfaces. A search engine can return a library record next to a government report, while an AI tool can synthesize information drawn from several kinds of sources into a single response. The access feels flattened even though the underlying information is still organized according to very different rules. When a search produces weak results, we may keep changing the keywords even though the larger problem is that we’re looking in the wrong place.

Before deciding what to search, then, we need to know what we’re trying to find. If we want to understand how the freeway route was approved, we need records of the decision-making process. If we want to know how the neighborhood’s population changed, we need measurements that allow us to compare the neighborhood before and after construction. Once we know what kind of information could answer each part of the question, we can ask who would have created or preserved that information and which research system is designed to help us find it.

What we need, then, is a way to see that larger research environment as a whole: the different kinds of places it contains, the roles those places play, and the paths that can carry a question from one to another. An atlas offers a model for doing that. An atlas can bring together different maps of the same territory, allowing us to see features and relationships that no single map can represent on its own.

From that model comes the Atlas of Learning, a framework for mapping the research environment itself.1 Instead of geographic territory, it maps where different kinds of information can be found, how those places and systems organize what they contain, and how we can move among them as a question develops. One part of the Atlas might show where original records are preserved, while another shows the systems through which published research can be discovered. Other parts can include places for structured learning or people whose expertise helps us find our way through systems we don’t yet know.

The routes among those parts are part of the Atlas too. A planning document may give us the formal name of an agency, while a scholarly article may give us a citation to an earlier source. An archival record may introduce the name of an organization whose records are preserved somewhere else. Those names, citations, institutions, and other clues allow one source to generate the next search.

The framework isn’t meant to contain every research resource that exists. It gives us a way to recognize different kinds of resources, understand what each can help us find, and see how one can lead toward another. As we discover new places and learn how to navigate them, we can add those destinations and routes to the Atlas ourselves.

Why we need the Atlas now

Before so much of research converged onto a common screen, the form and setting of a resource often revealed what kind of knowledge we were entering: a card catalog described and located materials, a statistical yearbook gathered institutional measurements, and an archival collection placed us among the records of a person or organization. Those distinctions can become harder to see when a government report, library record, and generated response arrive through the same interface. Many parts of the research world can now be reached from the same screen, but that ease of access can make those parts look more interchangeable than they are.

The Atlas restores that larger view. Instead of presenting every resource as another result arriving through the same interface, it makes the different parts of the research environment visible in relation to one another. We can see which places are built for which kinds of questions and how a discovery in one part of the Atlas can open a route into another. As more of our information environment is encountered through common interfaces or generated responses, that layered view gives us back a sense of the terrain we’re moving through.

The places and systems in the Atlas

The categories that follow are a starting set of destinations and discovery systems for the Atlas. They aren’t an inventory of every research resource in every field or country. They represent recurring functions that appear across many kinds of research: places that preserve original records or publish measurements, systems that help us discover books or articles, resources that teach us the language of a field, and people who can connect us to material we wouldn’t know to seek on our own. Once those functions become familiar, we have a way to approach a resource we haven’t used before: we can ask what it does, what it can help us find, and when it might be a good place to turn.

Catalogs and indexes

• What they are: Catalogs and indexes help us discover sources that may live somewhere else. A library catalog may tell us that a book exists and identify libraries that provide access, while a scholarly index may identify an article whose accessible copy lives in a university repository.

• What they’re good for: These systems give us a description of the source before we reach the source itself. Finding that record has already advanced the research because we now know what we’re looking for and where the material may be available.

• How to use them: Pay attention to the description around the source rather than treating the record as nothing more than a link. An author name may lead to another work, while an established subject term can reveal a broader group of materials. A catalog record can introduce a series title or an organization name we hadn’t known to search.

Archives and original records

• What they are: Archives preserve records created by people and institutions. A city planning office may have produced reports or correspondence over many years, while a community organization may have left meeting records or photographs. An archive often preserves or reconstructs relationships among materials associated with a person or institution because that context helps explain what the records are.

• What they’re good for: Archives belong in the Atlas when we need original records or earlier versions, especially material created close to the event we’re studying. They can take us closer to the records people and institutions created through their own work rather than later accounts of those records.

• How to use them: The way archival material is organized affects how we need to search. An archive may describe folders or boxes without making a separate searchable record for every document. The stronger route may begin with the institution that created the record or the name of the larger collection, then narrow once we understand how the material is arranged.

Data and public records

• What they are: Government agencies and research organizations collect data and records through their work, then publish some of those materials through data portals, reports, or records systems. Published statistics reflect decisions made before the number reached us, including what will be measured and which population will be included.

• What they’re good for: Data systems help us trace a number to its source, compare figures on compatible terms, or see how a measure changed over time. Definitions affect how categories are applied, while reporting periods determine which moments can be compared. By the time a number reaches a headline, those decisions may have disappeared even though they determine what the figure can tell us.

• How to use them: Return to the institution that produced the measurement and look at how the data was defined and collected. A city housing department may define a housing unit or program category differently from a federal statistical agency, while two employment series may cover different populations. Those distinctions affect what kinds of comparisons we can make.

Scholarly databases and research literature

• What they are: Scholarly research is distributed across journals, books, conference proceedings, repositories, and institutions, which makes dedicated databases and indexes valuable for finding work across a field. Those systems can organize papers by author or subject, while citation links show relationships among publications.

• What they’re good for: Instead of seeing one study at a time, we can begin to see the body of research surrounding a question. Scholarly databases help when we want to understand what researchers have studied or how an argument developed over time.

• How to use them: Treat one strong paper as an entrance into a larger conversation. Its reference list shows which earlier studies the authors relied on, while later papers that cite the study can show how researchers responded after publication. A review article can gather many studies around one question and introduce the terminology used across the field.

Guides and reference works

• What they are: Guides and reference works help us orient ourselves when a subject is still unfamiliar. A library research guide may identify the main resources used in a field and point toward relevant collections or institutions. An encyclopedia or subject reference work can supply formal names and established terminology that we wouldn’t have known to search on our own.

• What they’re good for: They help us learn enough about a subject to know what to look for and what language to use. They can introduce the major concepts used to describe a subject, the names associated with it, and the kinds of information that may be available to us.

• How to use them: Look for vocabulary and names that can sharpen the next stage of the research. A formal term can replace the everyday language we started with, while the name of an institution can give us another place to investigate. The goal is to leave with a better understanding of the subject and more informed options for where to go next.

Courses and structured learning

• What they are: Courses often arrange knowledge in sequence, while syllabi, lecture series, and open educational resources can provide a structured path into a subject. They become part of the Atlas when the problem has grown beyond finding one source and we need an organized way to understand the subject.

• What they’re good for: A series of searches can answer separate questions while leaving us without a framework for connecting the answers. Every explanation may introduce another concept we need to look up before the first one makes sense. A course can supply a framework that isolated searches may not.

• How to use them: Use structured learning to understand the relationships among concepts before returning to narrower research questions. A course on urban planning can introduce the ideas that connect transportation policy and neighborhood change before we encounter those ideas separately in reports or scholarly articles. A syllabus can also give us readings and terminology that make later independent research easier to follow.

People and institutions

• What they are: Librarians, archivists, curators, researchers, and other specialists can help us navigate parts of the research world that don’t connect neatly through databases or search engines. Their knowledge can include how collections relate to one another or how an institution describes and organizes its material.

• What they’re good for: A librarian may know which database covers a narrow subject, while an archivist may know that a person’s papers were deposited under the name of an organization. A curator may know that only part of a collection has been digitized, while someone who works in the field may know which professional organization maintains a hard-to-find resource.

• How to use them: The question doesn’t have to be “What’s the answer?” We can ask where to look and which institution is likely to have the material. We can also explain what we’ve already tried, which gives the person helping us enough context to suggest another approach.

Learning what these places and systems do gives us the landmarks, but an Atlas becomes more powerful when we can move among them. A research question can then move across categories as it develops. What we find in one place can change what we need to know, introduce a new destination, or give us language that makes another part of the research world searchable.

There isn’t a prescribed sequence among the different parts of the Atlas. Any one of these places or systems can potentially lead to any of the others, depending on what we find and what the question asks of us next. The routes develop through the research itself. Once we can recognize the different places in the Atlas and understand what each can offer, we can start deciding where a particular question should begin and how to follow the paths that emerge from there.

Finding a route through the Atlas

Knowing what kinds of places and systems exist gives us more options for where research can go. The harder question is deciding where to begin when we know what we want to understand but don’t yet know which resource can help us answer it. We shouldn’t have to know the name of the right archive, database, agency, or collection in advance. Figuring out where to look is part of the research itself.

The sequence below gives us an order of operations for getting started. Follow these steps in order when the research path isn’t yet clear. They’re designed to help us move from a broad question toward a plausible first destination without assuming we already know where the answer lives. Once we begin finding things, we can follow whatever new directions the research opens.

Break the larger question into things we can investigate separately.

Ask what we would need to learn before we could answer the larger question. If we want to know what happened to a neighborhood after a freeway was built through it, we might want to understand how the route was approved and how the neighborhood changed afterward. We may also want residents’ accounts from the period or later research about the consequences. Breaking the question apart reveals the different research tasks contained inside it.

Ask what kind of information would help answer each one.

We don’t need to know the source yet. We only need to describe what we’re looking for. The freeway approval calls for records of how the decision was made. Population change calls for measurements across time. Residents’ experiences call for accounts from people who lived through the project.

Use the Atlas to figure out where that information might live.

Official decisions may point toward public records or archives, while population measurements may point toward a government data source. If the likely destination isn’t obvious, guides, reference works, librarians, or other specialists can help us identify where to look.

Start with the strongest lead and pay attention to what it opens next.

A planning report may reveal the formal name of an agency we didn’t know to search. Meeting records may introduce a neighborhood organization whose records survive somewhere else. A scholarly article may give us terminology that exposes another body of research. What we find at the first destination changes what becomes possible at the next one.

A research route is emergent. We can make an informed choice about where to begin, but we can’t know the entire path before the research produces the names, terminology, institutions, sources, and connections that make later destinations visible. That also means getting redirected isn’t necessarily evidence that we started in the wrong place. A source that sends us somewhere we hadn’t anticipated has given us information we didn’t possess when we chose the first destination. The Atlas helps us recognize those openings and incorporate them into the path as it develops.

When we feel stuck or lost

Research can become disorienting after we’ve been following sources for a while. Sometimes we still know exactly what we’re trying to find, but the route we’re following stops producing anything that helps us get there. At other times, the research opens so many new directions, or changes our understanding of the question enough, that we no longer know which direction to follow. The first experience is being stuck; the second is being lost. They can feel similar because both leave us unsure what to do next, but the source of that uncertainty is different.

Being stuck means we still know what we’re trying to find, but the route we’re following has stopped helping us get there. A collection may contain less than we expected, while a promising citation may lead nowhere accessible. The terminology we’re using may also be a poor match for the way the relevant institution describes its material. In that situation, the question still gives us direction. The problem lies somewhere between that question and the route we’ve chosen to pursue it.

A different way into the same material may reopen the research. The kind of information we’re seeking can point toward another part of the Atlas, while a guide or a person familiar with the subject can reveal terminology or institutions we hadn’t encountered. Knowing that we’re stuck helps us focus on whether another route toward the same question is available.

Being lost means the research has generated more possible directions than we can follow at once. Each path may be legitimate. A planning record might open questions about who influenced the route, while census data raises questions about displacement and newspaper coverage reveals how residents responded. All of those directions could deepen our understanding of what happened to the neighborhood, which is precisely why choosing among them can become disorienting.

What restores that sense of direction is getting clearer about which part of the larger question we’re trying to answer now. If we’re investigating how residents were affected after construction, demographic changes have a more immediate role in that inquiry than the political negotiations that determined the route. The political history remains relevant to the larger subject, but it answers a different part of it. Seeing what each path can help us understand gives us a basis for choosing where to go next.

Being stuck and being lost therefore call for different kinds of reorientation. When we’re stuck, we still know where we want to go, so the next best step is finding another way toward that destination. When we’re lost, too many routes have become possible, so the next best step is identifying which one contributes directly to the question we’re trying to answer now. The Atlas can keep expanding without every path needing to be followed at once.

Drawing an Atlas of our own

An Atlas of Learning can begin whenever there is something we want to understand beyond a quick factual lookup. Each inquiry creates its own terrain because the places we need to visit depend on the question. A health question may take us through medical literature and public health agencies, while a historical question may lead toward archives, newspapers, or census records. The Atlas gives that particular inquiry a visible research environment of its own.

Because the Atlas is organized around a question, that question becomes the compass. It gives the research direction as new possibilities appear, helping us decide whether a newly discovered resource is relevant to what we’re trying to understand and what it might contribute to an answer. And it becomes the point we return to whenever the terrain grows more complicated, helping us recover our bearing and decide what to pursue next.

The working Atlas can be extremely lightweight. A spreadsheet, Notion page, mind map, document, or another information-management tool can all serve the same purpose. At minimum, we need the question, the research destinations that become relevant to answering it, and a brief note about what each can contribute. The form matters less than being able to look at the Atlas and see, at a glance, where this particular inquiry can go.

The Atlas represents the research terrain for a single question. It keeps the destinations connected to that question visible together, giving us a way to see the terrain as it develops and maintain our orientation within it. Something becomes part of the Atlas when it materially advances the inquiry, either by helping us answer the question or by opening a direction we may need to follow next. The result is a record of the sources and destinations that have become consequential to the research, rather than everything we happened to encounter along the way.

A larger world comes into view

Building an Atlas gives us practice locating a question within a larger information environment and recognizing how different parts of that environment can help answer it. We also learn to follow connections as new territory appears and to find our bearings when the direction becomes uncertain. Taken together, those abilities form a kind of information literacy we can apply when we enter subjects we have never explored before.

With that practice, entering an unfamiliar subject becomes an act of orientation as much as discovery. We may begin with little sense of how knowledge about that subject is organized or where its records, research, and expertise reside, but we know that those structures are there to be found. As they come into view, the subject begins to acquire a geography of its own, one that we can learn to read even though we have never traveled through it before.

The Atlas of Learning is something we draw as each question opens a new research terrain around us. The destinations will change from one subject to the next, and the route will develop differently each time. What remains is the ability to look beyond the interface in front of us and see a larger world of places we can learn to navigate.

You might also like:

• What “Do Your Own Research” Actually Means: A step-by-step methodology for evaluating any claim and exploring any subject.

• 27 Ways to Access Scientific Research: A complete guide to finding, reading, and evaluating scientific papers — and knowing what questions matter before you trust the findings.

• In Praise of Serendipity: On the joy of finding what you weren’t looking for, and how to get lost on purpose.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides.

Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

The impulse behind these archives isn’t new. In the first century AD, Seneca complained about Romans who amassed more scrolls than they could digest, warning that a wandering appetite for books left the mind scattered instead of nourished. A Meiji-era Japanese word, “tsundoku,” names the practice of acquiring books that pile up unread, a behavior common enough to earn its own word. The wish to keep more than we can absorb is ancient, and for most of history the collecting slowed itself: a new book cost money and claimed a place on a shelf we could watch fill up. Those prices made us choosy, so we brought home only what we expected to read, and the choosing was itself the curation: every book on the shelf had already survived a decision about whether it belonged there.

What’s changed is that nothing slows the collecting anymore. Saving costs nothing, takes no space we can see, and happens in the time it takes to move a thumb—so scale became unlimited at the same moment the friction that limited it disappeared. Nothing about the resulting pile reveals a new weakness in us. The old wish is simply operating in an environment engineered to strip away every limit that once contained the impulse.

The save button began with a promise. Read-later apps like Pocket and Instapaper appeared in the late 2000s offering a trade: tap save now, and come back to the piece when time allows. The button then spread far past the apps, to almost every content surface, from the bookmarks bar to the Notion page. As it spread, the design incentive shifted from serving the reading to serving the engagement metric, where each tap counts whether or not it leads anywhere. To us, the tap still feels like a plan: this piece, but later. The platform counting that tap reads it as engagement, one more sign the feed kept us interested, whether or not we ever actually return to the piece. The tap has become its own emotional payoff, and the queue behind it is where reading goes to be indefinitely postponed. Every part of the pipeline stays functional except the reading itself, which was the pipeline’s original purpose.

Living with the unread

The archive that grows this way is a record of single moments of attention with nothing connecting them: a tax explainer beside a poem beside a friend’s apartment listing, each saved on a different day for a different reason. Every save is a small message to a future version of ourselves, an ideal reader with the time and attention to sit with what we’ve caught. But the trouble is that the recipient keeps changing. The queue accumulates work addressed to a person who was about to become a parent, or start a business that never happened, and that reader never arrives to collect. Meanwhile each new save asks less of us than the one before, because past a certain size the pile can no longer be seen in one view, and adding to what can’t be seen carries no visible cost.

Left alone, the arrangement settles into a strange steady state. We keep encountering writing all day and keep feeling like readers, while the actual reading thins to headlines and opening lines, because the queue absorbs everything that would have taken longer. Opinions start resting on arguments we secured and never absorbed, and the sense of being informed floats free of anything we could explain to a friend across a table. The queue itself turns from a tool into a mountain to climb, since opening the bookmarks folder means facing every promise inside. So we stop opening the folder, and the place we built to keep what we valued becomes the place we avoid.

There’s a reason the save feels satisfying in a way that reading in the moment doesn’t, and it starts with what the brain does when it expects future access to information. In 2011, the psychologist Betsy Sparrow and her colleagues published experiments in Science that isolated the effect. Participants typed facts into a computer; half believed the file would be saved, half believed it would be erased, and nothing else varied between them. Those who believed the computer would keep the information remembered the information itself far less well, and remembered instead where it had been stored. The brain, sensing that an external store would keep the content, shifted toward remembering how to retrieve the information rather than the information itself.

That shift in memory, from the content to the route back to the content, is where the save’s satisfaction comes from. An unread piece we want is a small open demand, a claim on attention we can’t pay at the moment we meet it, but the tap to save closes the demand immediately: the piece is captured and safely stored, and nothing is pulling at us anymore. Reading would earn the same sense of completion only at the end of the piece; the tap collects it early and for free, before any reading has happened. Each save can deepen the habit, and the habit leaves a residue: an archive full of pieces we’ve never read but vaguely feel we know, because a filed address masquerades as familiarity with what the piece says.

Before the save button

Saving didn’t always outrun reading, because saving used to be labor. We folded down a magazine corner, tore an article out for a folder, printed pages at work to carry home, filed a browser bookmark into a folder we’d named ourselves. That labor did the curating: what made it into the folder had already earned a place there, and the folder stayed small enough that browsing it was a way of re-reading the collection. Each of those small actions also forced a moment of contact with the piece, a glance at what it argued and a judgment about whether it earned the effort. Saving started standing in for reading only when the effort disappeared and the two began to feel the same.

The disappearance came by degrees. Social bookmarking services had come and gone since the 1990s, but by the time Twitter added bookmarks in 2018 and every platform followed, the save button was a fixture on every surface with no shared sense of what saving was for. The evolution from manual to frictionless saving has transformed the act from a curated, high-effort task into an automatic, low-effort behavior, eliminating the cognitive appraisal required for content consumption. Deciding what to save was once a small act of judgment about whether a piece belonged in our lives, and the friction was what forced the judgment. Saving is now one tap that asks nothing, so when the friction disappeared, the judgment disappeared with it.

Working through the pile

Every item in the pile entered because something in it mattered to us at the moment of saving. Then the moment moved: the dinner that called for the recipe got rescheduled, the commute that would have carried the essay turned into a phone call. The pile is what those changed plans left behind, so its size simply measures how often life shifted. That’s also why actually working through it helps: a sorted pile returns the queue to service, and opening the bookmarks folder stops feeling like facing a cliff and starts feeling like browsing a shelf we intentionally chose. What remains is the practical question of what to do with the pile now; two workable paths answer that question, one fast and one more deliberate.

One distinction makes the work lighter before either path begins: some of what we save is an archive, and some of it is a queue. An archive is reference material, the tax resource or the recipe collection kept in case we want it someday; nothing in it is waiting to be read immediately, and a search bar is all it takes to find a piece again when a question comes up. A queue is promised reading, the articles and essays we told ourselves we’d get to; everything in it is waiting to be read, and time is all it takes for the waiting to turn into guilt. Most of the pile's weight comes from the two blurring, when reference material sits in the reading queue gathering guilt and reading promises sink into an archive where no one will meet them. Sorting out the blur starts with one question per item: is this here for looking up someday, or did I mean to read it as soon as possible? With that question in hand, the clearing itself can begin, and it can go one of two ways.

Path 1: The fresh start

Moving the entire existing archive into a single dated folder, out of sight but not deleted, lets us begin again with an empty queue: bookmarks drag into a folder named “Saved before 2026,” a Notion archive page gathers what came before, the screenshots slide into one album. The move borrows from email, where people whose inboxes grew past answering have archived everything at once and started over, a practice its survivors call declaring email bankruptcy. It works for the same reason here: nothing is destroyed and anything needed later can still be found by search, so the ambient weight of the backlog lifts immediately. The old saves become what archives have always been, a record kept for reference, ready on the rare day a question sends us back to look.

Path 2: The sorting pass

For those of us who’d rather look at what we’ve kept, everything runs on a single question, asked of each item: would I save this again today? A yes means the piece still belongs to the person we are now, and the follow-up is to read it soon or attach it to the work it serves. When the answer is no, the piece belonged to someone we’ve since stopped being, and releasing it costs nothing we still have. The question even settles the ambiguous cases: a three-year-old bookmark on Slack’s notification design goes, since newer writing exists and the curiosity that saved it has moved on, while an old Pinterest board from a finished renovation keeps its few strong images and sheds the rest. What survives stays small and current, which is what makes the remaining resources readable at all.

The fear of losing knowledge

Releasing an item retires a promise made by an earlier version of us who thought they’d come back and read the piece. Letting go can feel like losing knowledge, and the feeling traces back to the same mechanism: because the brain filed the location as a form of knowing, deleting the location reads as forgetting something we felt we had. The feeling misleads us, though. Since the reading never happened, the knowledge never made it into us; deleting the bookmark removes a route to information we hadn’t absorbed anyway. Remember: a piece we end up wanting later can be found again through search. Letting the item go releases that earlier self from the commitment, and the archive that remains starts to reflect what we’re paying attention to now.

A field guide to the moment of saving

Either path ends the same way: with a cleared pile. But a cleared pile refills unless the act of saving itself changes. The reflex fires the same way every time, but the situations it fires in keep changing: short on time, unable to focus, in the middle of work, deep in a feed. Each one calls for a different way of handling the piece we’re about to save. A single generic rule can’t reach across all of them, which is why “just have better judgment” reads as very unhelpful advice. Every fix below can run on effort in the moment, or it can be built into our tools so it runs on its own.

When we’re short on time but want to read this

The piece looks like a good use of time we don’t have right now, and the save feels like a fair compromise between wanting to absorb the piece and being unable to read it yet.

What helps

Reading the first paragraph before saving settles most of these. The opening often delivers the thing the headline promised, and the curiosity resolves on the spot; when it doesn't, the piece has earned its save, and the save gets a note about when the reading will happen, "for the train home" or "for the research on Thursday." A save with a schedule attached is a save that gets opened. Most tools have a field ready to keep it, since bookmarks take a description and a screenshot can be renamed.

Why this works

The bare save fails because the tap delivers the completion feeling before any reading happens, so the curiosity feels handled and moves on. Reading the first paragraph returns the piece's substance to the exchange, which is the difference between a save that rejoins the pile and a save that keeps its appointment.

When we can’t concentrate right now

The desire to read is real, but our attention is scattered. The save becomes a promise to give the piece the focused reading we don’t have in us at the moment.

What helps

A save made in a scattered state defers the same scattered state, since the future self who opens the queue inherits whatever attention we have then. If the piece belongs in our reading, it goes where focused reading happens for us, which for many of us is a particular place and ritual: the morning coffee before the phone, or the printout by the reading chair. Attaching the piece to that setting, and not to a general pile, gives the reading its best chance. If it doesn't rise to that, letting it go costs nothing, because a piece we couldn't concentrate on today was never going to fare better inside a queue of four hundred.

Why this works

Scattered attention is where the pile grows fastest, because a mind that can't settle saves without discriminating, and each save still feels like progress even though a scattered evening produces no reading to back the feeling up. Sorting these saves toward a reading setting, or letting them go outright, keeps the low-attention hours from building the archive we'll resent later.

When we’re working and find an article we like

Mid-task, an article surfaces that looks promising or maybe relevant to the job, and saving it feels like capturing something without breaking focus.

What helps

One question sorts these: what is this piece for? A piece that helps the current task gets used immediately, its one relevant point typed into the document we're already working in, and then nothing has to join any pile. Anything meant for a future project goes into that project's folder, where we'll see it when we next open the project. And whatever fits neither is a distraction, however promising it looks; closing the tab returns us to the work.

Why this works

A work-adjacent save is the hardest kind to refuse because it wears the costume of productivity while doing the opposite. The reach for the save button is a micro-break that fragments the attention the work needed, and the tap rewards the break with a small hit of accomplishment unrelated to either the task or the article. Asking the question takes the costume off: a piece with a job keeps its place in the work, and a piece without one stops looking like work at all.

When a roundup or a feed keeps offering more

A newsletter arrives with ten links, or the Instagram and Substack feeds stack recommendation on recommendation, and the saves come in runs: half a roundup banked in a minute, a dozen bookmarks from one night of scrolling.

What helps

Reading one link and leaving the rest settles a roundup. The email isn't going anywhere: it sits in the inbox, searchable by sender, and the other nine links will still be there on the day we want one of them, so saving them one by one just copies the newsletter into our pile. On Instagram and Substack, the in-app save is where posts go to be forgotten, so a recommendation that stays with us goes into the notes app or gets a when attached, and everything else scrolls on by.

Why this works

Feeds are built to surface more appealing pieces than anyone could read, so the sense of falling behind is part of the design. Saving relieves that feeling on the spot, which is why the saves come in runs: each one lets us pass a good piece without the small ache of losing it, and the scrolling continues. Reading one piece per session, or none, steps out of the arrangement, since the feed's output was never meant to be kept up with.

When a friend sends us something

A friend texts us an article with “you have to read this.” Saving it feels like doing something with the recommendation, a way of taking it seriously before we’ve read a word.

What helps

The friend wanted the reading and the conversation. Reading it now serves both when the piece is short; when it isn't, telling them when we'll get to it keeps the exchange alive, and a reply about the one line that struck us counts for more than a silent save ever could. Each of these keeps faith with the sharing in a way the untouched bookmark does not.

Why this works

A save is tempting here because it discharges the feeling of social obligation without the reading the obligation was about: the piece goes into the queue and the debt feels paid, while the friend hears nothing back. These saves pile up faster than most, because a link from a friend is harder to dismiss than one from a feed. But when saving becomes our standing response to them, the recommendations collect unread, and the conversations they were meant to start never happen.

When we save anyway

This is the case the others don’t cover. We can understand everything above and still watch our thumb reach for the save, because the reflex runs on the relief of the moment and never consults our track record. And the track record says most of these saves won’t be opened again. The fixes above all ask for a choice at the moment of saving, about what the piece is for and where it goes; here, no choice is coming.

What helps

The fix stops depending on the save entirely, and builds a return to the source. A save vanishes only when nothing ever brings it back, so the outcome turns on scheduling that return: a standing appointment, on the calendar, when the queue gets opened on purpose. What lasts is small and repeatable, something like one evening where we open the queue to read a single piece and release ten more without guilt. The appointment puts a floor of actual reading under the pile and keeps the pile small enough to stay in view, and what we save between appointments counts for far less once the appointment exists. Here, saving was never the problem; it was missing the return to what was saved.

Why this works

A habit doesn't stop because we understand it, and research on habit formation suggests why: behavior follows cues and repetition more than intention. An appointment moves the point of intervention off the save, where willpower loses, and onto a recurring block of time we set once and uphold like any other standing commitment. The pile stops being a place things disappear into and becomes a place we visit; a queue we visit is a queue that gets read.

After the pile

Reading starts happening again. The save marked "for the train home" gets opened on the train, the standing appointment works through the rest, and the reader who was always about to read actually starts reading. With the reading back, the background guilt thins out, since a queue we visit on schedule stops feeling like a debt. What the save button had been spending, without our noticing, was attention kept on one piece long enough for the piece to change us. That expense stays invisible at each tap, so the gap between what we saved and what we absorbed goes unnoticed until something forces a full look, a dying app demanding we export our saves, or a folder finally too full to ignore.

Evaluating a source has long stood at the center of information literacy, and the evaluation runs on two questions: who wrote this, and what do they want us to believe. Those questions have always been about reading a source’s intentions, which is why they work just as well one layer down, on the save button itself: who built this button, and what do they want from us? The answer isn’t hard to find: social platforms make money from the time we spend scrolling, and reading a long article means leaving the scroll, so their save buttons are built to file the piece in one tap and return us to the feed. Asking those questions gets us the same thing in both cases: a moment to decide for ourselves, instead of doing whatever the button was built to make easy.

Seeing the mechanism is what makes it possible to interrupt. The save button will keep appearing on every surface, engineered to be pressed, and the pull toward the tap will keep arriving with its small false sense of completion. All of it can keep happening while the reading comes back, because the change never depended on resisting the tap, where the reflex wins. The change comes from the return, the standing time when the queue opens and something in it gets read, so that saving stops being the place things disappear. Our archive was never a failure of discipline. It records a system working exactly as designed, and seeing the design is the first thing that has ever let us step outside its logic.

You might also like:

• A 30-Day Protocol to Rebuild Reading Capacity: Progressive overload applied to the page, for readers who miss reading.

• How One Rescue Mission Turned into the Global Record of Human Heritage: Inside the preservation system that classifies, catalogs, and protects humanity’s cultural and natural inheritance.

• The Libraries That Weren’t Supposed to Exist: Basement libraries and buried archives: a brief history of reading against the state.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides.

Upgrade to paid for the full Card Catalog. I appreciate you being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

This week:

• For twenty years, Google’s job was to send us somewhere else. New numbers show how fast that era is closing, and the publishers who live on those clicks are fighting for a new arrangement.

• A reporter spent months wearing AI devices built to watch and listen to daily life. The strangest moment arrived when one of the devices reached a conclusion about her inner life that she knew to be false.

• One of the year’s best-selling romance novels has attracted a second kind of attention, this time from researchers. What they found inside its pages raises a question about every book we pick up.

AI Answers Are Becoming the Way We Search

What Happened

The market intelligence firm Similarweb released an analysis of the generative AI landscape showing that Google’s AI Overviews, the machine-written answers that appear above traditional search results, now show up in 43 percent of searches in the United States. One year earlier, the figure stood at 15 percent. Visits to AI Mode, Google’s conversational search product, more than doubled in under a year, reaching 279 million by May 2026.

The analysis also found that we are adapting to the new interface. The average length of a Google search has grown over the past year, which Similarweb reads as users trading short keyword queries for longer, conversational questions written with an AI in mind. Searching increasingly begins and continues inside Google’s own products, with the AI answer sitting where a list of outbound links once began the journey.

Clicks out to the rest of the web tell a more complicated story. Citations inside AI answers grew more than fivefold over the past year, though only about 7 percent of ChatGPT desktop queries in the United States included them as of May 2026. After OpenAI made outside links more prominent in a May update to ChatGPT’s search experience, the share of those American desktop visits continuing on to an external webpage rose from 25 percent in March to nearly 60 percent by the end of May.

Why It Matters

For most of the web’s history, a search engine existed to send people elsewhere. The Similarweb numbers describe that role giving way to a different one, in which Google composes the answer from the sites in its index and keeps readers on its own pages. Much of online publishing runs on the advertising and subscription revenue that outbound clicks generate, and news publishers in particular have reported steep losses in referral traffic as the summaries have spread.

The web’s infrastructure has begun to respond. Cloudflare, an internet infrastructure company, introduced tools that let publishers block AI crawlers unless the companies behind them pay for access through a content marketplace. A negotiation is now underway between the companies that build AI models and the publishers whose work feeds them, and its outcome will shape what future answers can draw on.

What It Means for Us

The most immediate change is what greets us when we ask a question. By Similarweb’s measure, more than four of every ten Google searches in the United States now open with a synthesis written by a machine, with the sources tucked behind links we may never open. Leaving the summary to read the pages underneath has become the step that tells us whether an answer reflects what its sources say.

Our own habits are part of the data. Longer, conversational queries hand the system more context and draw back answers that sound more complete, which can make clicking through feel unnecessary. The ChatGPT referral numbers point in a hopeful direction: when outside links became easier to see, far more of us followed them, which suggests the appetite for sources survives the convenience of the summary.

Bottom Line: Interface design has become the lever that decides where reading happens online. The share of ChatGPT visits reaching outside pages more than doubled after a single update made links easier to see, and Google could move traffic on a far larger scale with choices about layout alone. The payment layer forming underneath is also significant: if marketplaces like Cloudflare’s gain popularity, access deals between AI companies and publishers could start determining which sources appear inside answers at all. The open question then moves ahead of the click, to which voices were available to the machine before we ever asked.

AI Gadgets That See and Hear Are Moving Into Daily Life

Read more

That saved transcript is what makes health and money the hardest ground to cross with an assistant. The very details that make the answer precise, an account number or a lab report with a name across the top, are the ones we have the most reason to guard. The good news is that the AI assistant doesn’t need them. A clear answer runs on the facts of a situation, not on who we are. We can hand over what the problem is, and keep everything that identifies us to ourselves.

Reader Question

“My father had a heart attack in March. Since then I’ve been pasting his test results into ChatGPT and asking what they mean, because his cardiologist talks fast and the appointments are short. Last month I started asking about money too: whether we should refinance the house to cover what insurance didn’t, and whether my retirement account can absorb the hit. The answers are clearer than anything I find by searching. Then a friend told me I shouldn’t be putting any of this into a chatbot. Is she right, and if she is, what am I supposed to do instead?”

— D.M.

A note before we begin: Thank you for being here, whether you’re a free or paid subscriber. The opening covers what AI is good and bad at when the subject is our health or our money, and what happens to the private details we type into it, all free to read. The paid section is the hands-on toolkit: the exact steps for sharing sensitive details safely, what a safe request looks like for a health question and a money question, and the four settings that decide what happens to a conversation after it ends.

Librarian Answer

The exposure people worry about concentrates in one gesture: the upload. A lab report or a bank statement is a crowded page, where the figures we came for sit beside a name and a record number that have nothing to do with the answer. Send the page, and every part of it lands in a transcript we no longer control. The AI assistant reads the two numbers we came for, but the platform may keep or review the rest

The way through is to change what the assistant is given to read, and the lever for that is the difference between a situation and an identity. An answer runs on the situation. A cholesterol reading and a date can take a health question as far as the full lab report does, and a balance with an interest rate can take a money question as far as the whole statement does. Typing those few figures into the chat instead of uploading the document leaves the answer whole and keeps the parts that name us out of the record.

Warnings about this usually stop at the phrase “sensitive information,” which leaves us to work out on our own what sensitive means. But the sorting is easier than that. Most of what sits on a statement or a test result does no work at all in answering our query, from the letterhead down to the transactions we aren’t asking about. Strip that away along with the name and the account number, and what remains is the handful of facts the question turns on.

What AI is good for

Money and health reach us wrapped in the private languages of their professions. Translating that language is what an assistant does best. It can explain what an A1C measures, or what an expense ratio does to a retirement account over the years. It can take a specialist’s dense paragraph and hand it back in words we could repeat to a worried parent at the kitchen table. Language built to be professional for practitioners becomes language the rest of us can carry into a decision. Few appointments ever run long enough for that translation to happen in the room.

AI is strong at preparation too. Once the terms make sense, it can draw up what to bring up at a cardiology follow-up or a mortgage meeting, more thoroughly than anyone tends to manage alone. It will lay out the trade-offs between a fixed and a variable rate, or between a generic and a brand-name drug, in the general form we want in hand before we sit down with the person answering for these particulars. Even a rehearsal conversation is helpful, whether we are practicing how to challenge a hospital bill or how to press an advisor on what their fees buy.

What AI gets wrong

An AI assistant answers with the same confidence whether we’ve given it the whole picture or half of one, and nothing in the reply signals which. A clinician spends much of a visit drawing out what we didn’t know to mention, because the detail that changes a diagnosis is often the one we assumed was beside the point. A chatbot takes our framing as given; a well-organized, confident answer can still rest on a version of events missing the fact that mattered most.

Agreement is the next problem. Handed a plan we have clearly already settled on, these systems lean toward endorsing it, a documented habit researchers call sycophancy. A cardiologist can tell us outright when a supplement is doing nothing, while an assistant nudged toward a hoped-for answer often reaches for something encouraging instead. Arithmetic is a separate weak spot, especially in multi-step calculations or when the numbers sit inside prose. And unless a model is looking things up as it answers, it works from training data that may predate the current interest rates and treatment guidelines.

Underneath all of that sits accountability, the question of who answers when guidance turns out to be wrong. A physician answers to a licensing board. In many countries a financial advisor can be legally bound to put a client’s interest first, with regulators and courts standing behind that promise. A chatbot operates under terms of service, which create no comparable duty. When its guidance goes wrong, the loss lands on us and nowhere else.

What happens to what we type in

A conversation with an assistant leaves a record behind. OpenAI’s chief executive has publicly noted that people often share with ChatGPT the kinds of things they would only tell a therapist or a lawyer, and that, unlike those conversations, chats with AI systems do not carry legal privilege. Privilege is the legal protection that can keep a lawyer’s or a doctor’s notes out of a courtroom; a chat has no such protection. In 2025, a federal court ordered OpenAI to preserve user conversations, including ones users believed they had deleted, as potential evidence in a copyright case.

A diagnosis spoken aloud in a clinic sits behind dedicated privacy law across much of the world, while the same diagnosis typed into a chat sits behind a privacy policy the company can rewrite whenever it likes. Under some default settings, these conversations may be used to improve models, or reviewed by human evaluators checking system quality. For anything touching our health or our money, that means our own words could be folded into a future model or surface in a legal case, without our ever knowing it happened.

The identifiers we strip out are the exact parts that let someone act as us or against us. A name beside an account number is the raw material of fraud. A diagnosis tied to our identity can change what an insurer offers or what an employer assumes. Once any of that sits in a transcript a court can reach, a private moment can resurface in a divorce or an insurance claim, long after we have forgotten the conversation. Stripping the identity out is how we keep a passing worry from hardening into a permanent, transferable fact about us.

All of this is manageable at the point of entering in the chat box. A question written without a name or an account number stays ours no matter where the transcript travels, and it comes back with the same answer. Information literacy usually means reading a source carefully. Here it also means reading the tool, then deciding what to withhold.

How to share the sensitive parts safely

Read more

That exchange violates a rule of information etiquette. Information etiquette is the interpersonal counterpart to information literacy: the layer of unwritten social norms governing how we handle information with each other, alongside the individual work of evaluating what’s true. Its rules are what we feel getting broken when a message lands wrong, invisible until somebody breaks one and unmistakable in the moment they do.

Small breaches like this one populate the daily texture of our messages. Screenshots arrive in family text threads ripped from contexts we can no longer reconstruct. A colleague drops a link into Slack with a casual “thoughts?” that reads to everyone in the channel as a pointed attack in disguise. A friend forwards a scientific study to prove they were right about something that was discussed three weeks earlier. Somebody shares a piece we wrote and misrepresents it in the caption, then tags us so the distortion lands directly in our notifications.

Underneath each of these moments is a rule being broken. We feel these rules in the flicker of hesitation before we hit send, and in the discomfort when a message arrives that we didn't want to receive. They register in gut reactions we can’t always articulate to ourselves.

We can feel a violation without being able to name the rule that was violated, because nobody ever taught us the rules. Information literacy has been the focus of educational attention for decades, but that attention has been directed primarily at how we consume information. How we treat each other during the exchange itself has received comparatively little focused attention, and the vocabulary for talking about it remains fragmented and underdeveloped.

Etiquette exists in nearly every other domain of human contact, which is what makes its absence from information exchange feel so sharp. Table manners govern how we eat together, and Emily Post built an influential twentieth-century empire on the premise that shared expectations reduce friction and build trust between people who might otherwise misread each other. Journalism and academia developed their own attribution standards over generations, though both were designed for professional settings that don’t map onto the informal channels where much of our contact with each other now takes place.

Information etiquette isn’t a term in wide use, though some of the ground it covers has been mapped under other names. Virginia Shea’s 1994 book Netiquette named a set of rules for online conduct focused mostly on early Internet contexts, and digital etiquette now often shows up in educational and workplace training. Neither has consistently centered the interpersonal dynamics of how we handle information between people who already know each other, in the messaging environments where much of our informal contact now happens. The rules named here also live inside a particular cultural context of contemporary American and Western digital communication, where directness and individual attribution are baseline expectations that vary substantially in other parts of the world.

The rules of information etiquette can be traced by looking at where they get broken. Six come up over and over across every kind of exchange between people who know each other. Each is a norm we already follow most of the time, and every one registers as felt friction the moment it gets broken.

The six rules of information etiquette

Each of the rules that follow works the same way. The sender does something small before the message goes out, like reading the article first or adding a line of framing, and the recipient feels the difference when that care is missing. Leave out the small step often enough, and the recipient starts trusting our messages less. The rules are recognizable from both directions: we've each been the sender who rushed past them, and we've each been the recipient who noticed.

We recognize these violations long before we have words for them. Most of us can recall the small drop of receiving a link that was clearly sent to score a point, or the low-grade irritation of realizing a sender never read what they shared, well before we could say what norm had been crossed. The entries below supply the vocabulary. Each one names the practice and describes what its absence feels like from the receiving end, so that both sides of every exchange become visible at once.

The moment to employ each rule is the moment before sending. Every one of the six condenses into a question that can be asked in the second before a message goes out, listed beneath each rule below.

Rule #1: Read before we share

Ask before sending: Have we read this?

When we forward a piece we haven’t read, we’re asking somebody else to invest attention in material we haven’t invested our own attention in. Sharing implies a basic endorsement, or at minimum a statement that the content is going to be useful for the recipient. The unread share tends to reveal itself: the recipient opens the piece and finds contents that don’t match how the sender framed them, or brings up the article later and discovers the sender has nothing to say past the headline. Each discovery accumulates, and over time recipients start taking that sender’s shares less seriously.

The tracking often happens below the level of conscious awareness, in the same way we register who runs late and who follows through on plans. Over time we come to know which of our people read what they send and which forward on headline alone. The habit is common enough to measure: a 2016 study by researchers at Columbia University and Inria, the French national research institute for digital science and technology, found that 59 percent of the links shared on Twitter in their dataset were never clicked. Even if the exact numbers have shifted with platform changes since, the pattern is recognizable from most people’s group chats and feeds.

Sharing without reading also changes what spreads. When a forward is a reaction to a headline, the conversation that follows is about what the headline implied instead of what the article said. The person who received the link and read the entire piece often knows more about what got shared than the person who did the sharing.

Rule #2: Don’t send an article in place of a conversation

Ask before sending: Why are we sending this?

Sharing sources is part of how disagreement gets resolved. When we lay out the reporting that changed our mind and invite a response, the article serves the conversation; the evidence gives the other person something concrete to answer. Arguments built on shared sources tend to go better than arguments built on assertion alone.

This rule covers a different send. The same article arrives bare, with no note about why we’re sending the piece and no opening for a reply, sometimes days after the argument went cold, sometimes dropped into a group chat where one particular person is expected to feel its sting. A source offered inside a conversation asks for a response. But a bare link announces instead that the sender considers the matter settled, a way of saying “I’m right” without saying the words.

Arguing this way means the argument never happens. The recipient feels the sting, the sender walks away believing the point got made, and the disagreement stays exactly where it was. A forwarded article can't do the work of a direct conversation, and every send that substitutes for that conversation makes the conversation harder to eventually have.

Rule #3: Warn people before sending heavy news

Ask before sending: Does this need a warning?

A friend texts a link to a long investigative piece about child trafficking with no framing about what’s inside. A coworker forwards a mass layoff rumor with a casual “fyi.” The emotional loads are radically different, but neither send reflects the difference in its framing; the recipient in each case has no way of knowing the depths of what they’re about to open.

Handing someone a piece of information means transferring an emotional experience along with the data. The recipient has to metabolize whatever the content contains on their own time, and often in whatever context they happen to be sitting in when the notification arrives: a doctor’s waiting room, the train home from a difficult day, the moment they were trying to fall asleep, the middle of a family dinner. The link doesn’t ask about their situation before it arrives.

A sentence of framing before a link changes how the link arrives. The framing can look like “this one’s rough, no need to open now” or a check-in about whether the topic is welcome before dropping the link at all. That small act gives the person on the other end a chance to choose when to open whatever’s inside, a chance they don’t get when a heavy link arrives bare.

Rule #4: Don’t ask for research and then reject the answer

Ask before asking: Are we open to the answer?

Someone asks for help finding information, whether a supplement is safe to take alongside a medication, or what to make of a legal situation they’ve been dealing with for months. The person asked does the work: they search, they read, they evaluate the sources, they synthesize what they find, and they bring back an answer. The person who asked then rejects the findings because the answer doesn’t match what they had already decided.

The person who did the searching finds out only when the findings get waved away: the ask sounded like a request for information while working underneath as a request for agreement. The asker usually isn’t being unfair on purpose: from the inside, wanting to be told we’re right can feel identical to wanting to know what’s true, especially when the answer we’re hoping for already feels obvious to us. Noticing which of the two requests we’re making before we ask, and receiving whatever answer comes back with an open mind, keeps the ask what the ask claimed to be.

The rule also gets broken from the other side of the exchange. Someone asks a straightforward question, and their response comes back as a curt “just Google that,” turning a genuine ask into a dismissal. But that response misses the social dimension of the original question: people often ask other people instead of searching because they want a human filter, somebody they trust to help them make sense of what they’re asking about.

Rule #5: Don’t share words without their context

Ask before sending: Is the context intact?

Screenshots are a common vehicle for this pattern across the platforms where information moves. A text message, a tweet, a headline, a data point, or a quote gets pulled from its original setting and dropped into a new one where it means something different from what it originally meant. The words themselves haven’t changed. What’s changed is the frame around them, and the frame was doing most of the work of establishing what the words meant in the first place.

Removing context produces different information that still looks like the original. The cropping can be careless, a screenshot grabbed in passing, or deliberate, when the stripped version says something the sender wants said. Either way, the person on the receiving end has no easy way of knowing what got removed, or how the removal changed the meaning of what they’re looking at. The distortion can become accepted as truth by people who receive it secondhand, and whoever wrote the original message may never get the chance to correct the record, or even know the stripped version is circulating.

Rule #6: Name the person who told us

Ask before repeating: Does the source have a name?

This rule applies to information that changes a decision or gets repeated onward. Most passing facts move through conversation without needing a name attached. But when information will steer a choice or travel onward, though, the source is part of the information: who stands behind a claim tells the listener how much to trust the claim and where to check further.

The habit stays light: saying “my friend who works in public health told me” or “I read this in a piece by so-and-so” instead of defaulting to “I heard somewhere that” or “studies show” or “apparently.” The named version tells the listener where the claim comes from, so they can judge for themselves how much trust the claim deserves. The anonymous version hands them a claim with nothing behind it, wearing the authority of research nobody can locate; the naming also keeps knowledge attached to the people who produce it.

How to repair a broken rule

Knowing the rules doesn't mean keeping them consistently all the time. We all forward without reading occasionally, and we all send heavy material without framing during moments of distraction. What counts as a violation depends on the particular two people involved: a link sent with no note could be routine between old friends but can land as rude between colleagues, and a "just Google that" from a manager lands harder than the same words from a peer. Cultural context moves the lines too, since directness that reads as considerate in one culture can come across as blunt or intrusive in another. The etiquette lives in noticing our own violations and other people's, and in what we choose to do in the aftermath.

Charitable interpretation is part of the noticing. A friend’s late-night forward with no context can be read as thoughtless, or as a gesture from somebody who was reaching out at the end of a long day. The same forward becomes a different message depending on which interpretation we choose, and we’re making that choice every time we open a notification. Interpretive charity means keeping the second reading alongside the first before deciding what a message meant, and staying open to the possibility that the person on the other end was working with different assumptions or moving too fast to notice what they’d just done.

Repair is the practice on the sending side. A brief acknowledgment that a send arrived at a bad time, or that framing got left off when the material was heavy, does the work of putting the violation into shared language. Something as simple as “sent that at a weird time; no need to respond tonight” or “sorry for dropping that without context, the topic caught me off guard” repositions the exchange without either party having to escalate.

Why these rules decide who trusts us

The rules operate one send at a time, and their effect accumulates across every exchange. A group chat becomes a place where hard news can be shared without wrecking anyone’s evening when the participants share basic norms about how sending works, and a family text thread becomes a place where questions can be asked without becoming a demonstration of who knows what. The norms make possible the trust that lets people stay in relationship with each other across the volume and speed of contemporary information exchange. Without them, the same exchanges become sources of low-grade anxiety about who’s going to send what next, and how it’s going to be framed.

There’s usually somebody in our lives whose handling of information registers as care. They pause before sending, frame heavy links, attach names to sources, and calibrate corrections in ways that leave the other person feeling seen. Their presence in our lives registers as relief, even when we haven’t yet named what exactly produces the relief.

That person isn't operating from any special knowledge; they just tend to pause at the moments the six rules describe. What changes when the rules get named, though, is that any of us can do the same noticing. The pause turns each send into a choice made with the recipient in mind. That consideration is what turns reflex into practice, and what turns a group chat, over months and years, into a place where the people we care about want to spend their attention.

You might also like:

• 20 Small Acts of Media Literacy We Can Practice Every Day: From checking the date to tracing the source: building a sharper eye for information.

• The information diet isn’t enough. What we need is an information village.: What changes when we go beyond optimizing our consumption and start building shared infrastructure.

• The Hierarchy of Sources: A Cheat Sheet: A guide to evaluating information sources in the AI age.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides.

Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

Reading at length is a physical and mental skill, built by practice and weakened by neglect—no different in that respect from the stamina to run a distance or carry a conversation in a second language. Years of reading in snatches, on screens, between the thousand small interruptions the screen invites, can train our attention toward speed and away from depth. The muscle for a long, unhurried text has gone slack from disuse, and the book that once carried us for an hour now loses us after a few pages. A slack muscle, though, is one we can train back to strength.

Part of what changed is physical, down at the level of the eye. Reading a printed line mostly relies on rapid eye movements called saccades, with brief pauses between them where the eye takes in the words before moving on. Scrolling often engages a different pattern: the eye can track text or other content sliding up the screen and follow it, the way it would track a bird across the sky, using what researchers call smooth pursuit. Heavy, habitual up-and-down scrolling likely gives that continuous tracking motion more practice, while the stop-and-go left-to-right rhythm that carries us through a printed page gets less exercise. What the evidence supports is the broader point: the difficulty sits largely in habits we’ve learned rather than in any fixed flaw within us. The feeds that trained us into scrolling were built by teams whose success is measured in how reliably they can pull our eyes back to the screen. Years spent inside that design leave their mark on how we read, and habits built by practice can be rebuilt the same way.

The screen-reading penalty has been measured directly. A 2018 meta-analysis in Educational Research Review pooled more than fifty studies comparing how much people understood when they read the same material on paper versus on a screen, across a large combined sample. On average, paper won, though the advantage was not uniform. It showed up clearly for informational reading, the kind we do to learn something, and grew larger when people were reading under time pressure. For narrative reading, the gap was much smaller and often negligible, which tracks with the feeling that a good novel can read just fine on a screen. One finding stands out: people reading on screens tended to think they had understood more than they truly had, so the screen hid part of its own cost from them. Screens remain superb for finding a fact or skimming a page. When the task is the slow work of following an argument that unfolds over many pages, paper reading tends to keep its edge.

Maryanne Wolf, a cognitive neuroscientist at UCLA who has spent decades studying how reading develops in the brain, points out something that changes how we see the whole problem: the human brain is not born with a dedicated reading circuit. Spoken language is old enough in our species that we come prepared for it from the start, and any child surrounded by talk will absorb speech without formal teaching. Written language is a much more “recent” invention by comparison. Each of us assembles the machinery for it through practice, wiring together regions that evolved for older purposes like seeing and speaking. The circuit they form didn’t exist in us before we learned to read. Wolf calls the fullest use of that circuit deep reading, the slow and immersed mode where we sink into a long argument and keep several pages of reasoning in mind at once. Because this circuit is built through practice, it remains elastic throughout life: it strengthens with sustained, demanding reading and can weaken when that kind of reading falls away. What slips is the stamina for a long and demanding passage, the part that took the most building in the first place.

But the same adaptability that lets the circuit weaken can also bring it back. It responds to whatever we practice, so the reading we’ve stopped doing rebuilds through reading we take up again. Nothing about the capacity has been permanently lost; it has atrophied from neglect, and it firms up again once we put it back to work. The one difficulty is that putting it back to work tends to fail when we rely on willpower and a free Sunday; what the rebuild needs is a method, not just motivation. The one that follows sets a pace, though capacity returns faster for some readers and slower for others. The schedule is meant to be adjusted to how our own reading responds, so it works best as a flexible guide rather than a rigid day‑by‑day rule.

Why the method works

The method that best fits this kind of rebuild comes from a field that seems far from books. In the mid-1940s, an army physician named Thomas DeLorme developed progressive resistance exercise programs for soldiers recovering from orthopedic injuries sustained in the Second World War. He found that weakened muscles rebuild most effectively when they are regularly exposed to loads that challenge them without causing injury: a load too light gives the body no reason to adapt, while one too heavy pushes it into strain rather than growth. Recovery happened in the narrow band between those failures, and the band moved upward as the muscle got stronger.

Reading endurance can be rebuilt on similar terms. When a session runs a little longer than our comfortable stretch, the reading brain extends itself to meet the demand. Anything within what already feels easy asks little of it, and a punishing session that ends in exhaustion only sours the next attempt. The zone that does the work sits just past comfortable, and it moves upward as our reading capacity grows.

Two things separate a rebuild that sticks from one that stalls, and the first is rest. Skills consolidate between practice sessions, and some of the neural changes that support new capacities occur during off-task periods and sleep, in the hours beyond the practice session itself. Short, repeatable sessions with space between them are more likely to build lasting endurance than one heroic afternoon, because much of the growth happens in the gaps. The second is restraint. The most common way a reading rebuild falls apart is the four-hour Sunday marathon that feels like triumph while it lasts—but leaves us avoiding the book for the next two weeks. That kind of afternoon builds nothing that lasts. It produces one memorable session and then a slow slide back to not reading, which is the exact outcome the protocol is designed to prevent.

What the protocol involves

The rebuild runs thirty days. Reading days alternate with rest days, and each reading session grows a little longer than the last one across four weeks, so that by the end we can sit with a book for a long stretch again. The alternating structure is what separates this from simply resolving to read more. Left to willpower alone, reading loses out to easier pulls on our attention; a fixed sequence of small sessions gives the capacity a dependable way to grow back.

The thinning of reading capacity shows up in two different ways, and the protocol comes in a version for each. For some of us, the endurance itself is gone: the words start to blur after a few pages, and a chapter we would once have moved through steadily now feels like a hurdle. Track one rebuilds that endurance. Others can still read for a long stretch when nothing interrupts them, but their focus fractures a different way. A common issue is when the hand drifts to the phone in the middle of a paragraph, moving on its own before we’ve consciously chosen to pick it up. Track two is for readers in that situation.

Choosing a track

The two tracks train different things: one rebuilds the endurance to read for a long stretch, the other strengthens the composure to stay put without reaching for the phone. Sorting ourselves correctly saves weeks of ineffective training: a reader with intact stamina gains little from lengthening sessions they can already sustain, while a reader whose attention has thinned needs that length rebuilt before the phone rule can take hold. The lists below set out the everyday experience of each, and the one that sounds like our own evenings points to the track to begin with.

Track one is the one to choose when the reading itself has become hard to sustain:

• The words blur or the mind drifts after a few pages

• What happened on page ten is gone by page twenty

• A chapter of a book we want to read starts to feel like a climb

• Books get abandoned partway through far more often than they get finished

Track two is the fit when the reading is fine but the focus keeps breaking:

• Reading for a long stretch is possible whenever nothing interrupts

• The hand goes to the phone in the middle of a paragraph, on its own, out of habit

• Lines get reread because a glance at the phone cost us the thread

• The book lies open, but the phone keeps drawing our eyes away from it

Many of us recognize ourselves in both lists at once, with thinned endurance and a restless hand working together. If that’s the case, we start with track one and rebuild the stamina first, since a longer attention span makes the phone easier to resist. Once we reach the second week, we fold in track two’s central rule, keeping the phone in another room during sessions, and let the two rebuilds run together from there.

Setting up: what to read

With a track chosen, a few small preparations decide whether the first session takes or fizzles, and the most important is the choice of book. A book that pulls us forward on its own does more for the rebuild than any amount of discipline. Staying on its pages costs far less willpower than wrenching ourselves through something we picked up out of obligation. A literary novel we’ve meant to read for years can do this, and so can a paperback thriller; the genre is beside the point next to whether the book makes us want to keep going. We choose the book we truly want rather than whatever we feel we ought to have finished by now, and we let the book’s own momentum carry part of the load.

The rest of the setup is light. A track one reader only needs the phone silenced during a session. Track two asks for one thing more: the phone waits in another room rather than face-down on the table. A phone within arm’s reach keeps the reaching reflex alive even while it sits untouched. To keep time without reaching for the phone, a kitchen timer or a wristwatch does the job, and a wall clock we glance at works just as well.

The last thing to settle is what to expect of the baseline reading number. If our reading has thinned, the first session may run only five or ten minutes before attention gives way, and that’s no verdict on us. It’s just the base we build from, the way a rebuilt strength program starts well under the weight it will finish lifting. A modest starting point followed faithfully for thirty days carries us further than an ambitious (or aspirational) one that collapses in the first week.

Setting up: paper, e-reader, or phone

The surface we read on is the one setup choice that earlier research speaks to directly. The answer is not the blanket rule that everything has to be printed on paper. For a novel or any reading where we mainly want to follow a story, studies generally find a much smaller or negligible difference between paper and screen comprehension compared with expository texts, especially when distractions are controlled. A phone reading app or an e-reader can be a fine home for track one, provided notifications are silenced and other competing apps stay out of reach.

Informational or study-style reading is a different case. This is the kind of reading where the paper advantage tends to be clearest, and where our tendency to overrate what we absorbed on a screen does the most harm. It’s also where rebuilt attention pays off most. A dense argument is precisely the sort of text a fractured attention skates across while feeling as though we actually followed it. For anything demanding, then, paper or a dedicated e-reader with electronic ink is the better choice. A phone reading app remains a good home for a novel, but it’s a poor place to attempt the harder work of rebuilding focus while every feed and alert that trained our distraction lives on the same device.

Running the protocol day to day

With the book and format chosen, what remains is a simple daily rhythm. Day 1 has one job, which is to find our starting number. With the phone handled according to our track, we sit down and read until the attention runs out: the point where the eyes are still moving across the words but nothing is registering, where we reach the foot of a page and could not say what it held. That is where the session ends. For track two, the session ends instead at the moment the pull toward the phone wins. Either way, we note the minutes, and that number becomes the starting length every later day builds from.

The even days are rest days, and skipping them is the surest way to stall. The gains from a reading day are laid down in the brain during the hours afterward, while we are away from the page, so the rest day is when the previous session turns into lasting capacity. Reading hard every single day can leave little room for that consolidating work, which is why the rest is built in. That said, a rest day is of course not a ban on books. If a rest day finds us wanting to read and free to do it, we can read gently, at or under the starting length. What the rest day guards against is added load; a few pages read for pleasure don’t offset the protocol.

The odd days from day three onward are reading days, each carrying a target length from our track’s table. We set a timer for the target, then read until the alarm sounds. If attention breaks before the timer does, that day’s target got ahead of us. We ease back to the previous length for the next reading day or two until it settles, then climb again when it feels ready. Falling short of a target happens often, and it means the pace needs a small adjustment rather than signaling that anything has gone wrong.

These increments come from applied principle rather than from a study built to test them. The study that would hand us a perfect minute-by-minute schedule for rebuilding adult reading stamina has never been run, a gap in the literature waiting for the right researcher to fill. The progressions in these worksheets are heuristics designed to apply the principle of progressive overload to everyday reading; they’re meant as adjustable starting points rather than rules validated in a trial. What decades of rehabilitation research have confirmed is the principle underneath: a weakened capacity rebuilds fastest when each session asks for a little more than the last one comfortably gave, and it stalls or breaks when the jump is too large. The additions in the worksheet are sized to stay inside that range for a common starting point, and they are meant to be adjusted to our individual levels.

Track one: rebuilding reading stamina

Track one usually starts from a low number, since endurance is the part that thinned. Five or ten minutes before the words blur is a common first reading, and that number is the base every later day builds from. The additions stay small at the start, and grow gently from there as the capacity returns.

The thirty-day protocol is easier to follow with something to write on. This printable worksheet gives track one a place to record the book we’ve chosen, our start date, and the starting number we measure on day one, alongside a log that already carries the target length for every reading day so the next step is never a calculation. Rest days are marked so they read as part of the plan, there’s room to note how the sessions feel as they shift, and a closing page of reflection prompts waits at the end of the month.

Card Catalog Classroom Reading Protocol Worksheet

1.42MB ∙ PDF file

Download

Download

The early additions stay small on purpose. A few minutes is a large share of a short session; three minutes on top of ten asks the reading brain for about a third more than it just gave, a stretch it can rise to. But ten minutes on top of that same ten would nearly double the session in a single day, and doubling is the overreach that backfires. Later in the month the same additions come more gently, because by then each one is a smaller slice of a session that already runs long. Day thirty carries no target at all. It’s an open session, a chance to feel what our attention will do now on its own, with no clock governing the reading.

Track two: rebuilding phone-free focus

Track two usually starts from a higher number, since endurance was never the trouble. Twenty or thirty phone-free minutes is a common first reading. If our starting number sits well above ten, we treat the table’s additions as the floor and stretch them wherever the sessions feel easy.

The track two worksheet works the same way, built for readers whose stamina is intact but whose focus keeps slipping toward the phone. It records the book we’ve chosen, our start date, and the phone-free number we measure on day one, with a log that fills in each session’s target as the phone-free stretch grows longer. Rest days are marked, blank lines wait for whatever we want to record about how the reach fades, and a short set of questions closes out the month.

Card Catalog Classroom Phone Free Focus Worksheet

1.42MB ∙ PDF file

Download

Download

What track two rebuilds is the fading of the phone-reach. Every session weakens the automatic pull toward the phone by a little, and the effect accumulates across the month. By the final week, an unbroken hour and then ninety minutes test whether reading has gone back to feeling like reading, rather than like a contest against the phone running the whole way through. If those last sessions still feel like a fight when day thirty arrives, we give the previous length another week before calling the protocol done.

When to adjust the pace

If the additions feel steep from the start, we halve them and let the rebuild take longer. For a first session of five minutes, even a three-minute jump is a big leap, and a two-minute one may sit closer to the band where the stretch does its work. Finishing a gentle climb does more for us than abandoning a brisk one somewhere in the first week. Also, nothing about the number thirty is sacred. Thirty days is a workable default rather than a finding, and a very low starting point or a thin daily margin may call for stretching the same steps across six or eight weeks instead. The steady climb in reading time is what rebuilds the capacity, however many days it ends up taking.

And life will interrupt, because it always does. And when it does, we simply pick up at the last day we completed rather than doubling up to catch up. A missed day is a pause rather than a loss, and the sequence resumes where it left off. Doubling up to erase the gap is the overreach pattern presenting itself again in so many ways.

Signs it’s working

The minutes climbing are the easiest measure of progress, and also the shallowest one. The changes to watch for are subtler and show up in how reading feels. Somewhere in the second or third week, the drop into a book tends to come faster. The first few pages stop being a warmup we have to push through, and the text takes us in almost as soon as we sit down. The pull toward the phone, for track two readers, goes from a constant pull to an occasional one we notice and set aside without much struggle. A third sign is slower to appear and the most telling: we start remembering more of what we read the day before, because attention that stays in place long enough is what lets a page settle into memory at all. None of these arrive on a fixed schedule, and noticing even one of them means the rebuild is under way.

After day 30

Thirty days in, readers might find that their renewed capacity to stay with a book for longer stretches feels self-sustaining, and the protocol scaffolding can come down. The alternating rest days can fall away and the timer goes back in a drawer. The phone can even return to the desk, though perhaps not all the way to the reading chair. What keeps the gains from slipping is simply continuing to use it, which for most of us means sustained reading on most days, in whatever stretch of time the day allows. The capacity behaves like the strength it resembles: worked in the ordinary course of a week, it maintains; abandoned entirely for months, it thins again, though rebuilding it a second time often comes faster than the first with the muscle memory in place.

The rebuilt capacity reaches well beyond the number of books we can finish in a year. Every day now arrives as a flood of claims competing for our attention and our trust, most of them dressed as confident summaries of things we have no time to check for ourselves. Critical thinking is what lets us tell the solid claims from the flimsy ones, and it depends on a skill we rarely connect directly to it: the ability to stay with a piece of writing long enough to weigh what it says and test its logic. A mind that can follow a long argument will follow it down to the assumptions it rests on, where the deepest disagreements usually hide. That same mind can carry two accounts of the same event at once, reading them against each other until the point where they divide comes into view. It can catch the paragraph that contradicts what an earlier one promised. None of this is possible without deep reading. Skimming takes the gist and moves on, which means it takes the summary someone else prepared and moves on, leaving us knowing whatever that person wanted us to know.

That capacity is the ground beneath the whole discipline of reading well. Weighing whether a source can be trusted, or tracing a claim back to the study it came from, depends first on being able to read the thing deeply at all. The judgment we bring to information rests on an attention that can stay in one place, and that attention is what these thirty days rebuild. A reading life that began as a private pleasure turns out to underwrite an independent mind, the difference between examining the world for ourselves and accepting the version we’re handed. That is the deeper measure of the rebuild. The stamina returns first, and the critical thinking that rides on it follows close behind.

You might also like:

• What Scrolling Did to Reading: Why books feel harder, what the science shows, and how to rebuild the capacity.

• Misinformation Spreads at the Speed of Impatience: Why critical thinking starts with slowing down, according to the research.

• Steel-Manning Can Be More Useful Than Fact-Checking: How engaging with the strongest version of an opposing argument builds a deeper defense than verification alone. A practical guide with a downloadable worksheet.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides.

Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

This week:

• Every prior intergovernmental AI body has included Western democracies among its members. A new one formed this week without a single Western signatory.

• Testing AI image models for whether they could generate illegal content has meant generating that content. A team of researchers just published a way around the problem.

• Economists have historically resisted claims that a new technology will transform society faster than institutions can adapt. A large, credentialed group of them just published exactly that claim about AI.

A New Global AI Governance Body Just Formed Without the West

What Happened

Twenty-nine countries signed an agreement in Shanghai this week to establish the World Artificial Intelligence Cooperation Organization, a new intergovernmental body headquartered in the city. Chinese Foreign Minister Wang Yi signed on behalf of China. The other founding members include Russia, Belarus, Serbia, Cuba, Brazil, Venezuela, Pakistan, Indonesia, Kazakhstan, and Laos, alongside additional signatories from across Africa and Asia. United Nations Secretary-General António Guterres attended the ceremony.

The signing came on the eve of the annual World Artificial Intelligence Conference in Shanghai, where Chinese President Xi Jinping is expected to lay out Beijing’s vision for the country’s role in global AI governance. China first proposed the organization at last year’s conference, and no country had formally committed until now. The United States, European Union, United Kingdom, and Japan are absent from the founding members’ list.

Why It Matters

Global AI governance is starting to take shape along two parallel intergovernmental tracks. The existing frameworks that carry weight in Western capitals, from the OECD’s Global Partnership on AI to the European Union AI Act, are anchored in the world’s leading AI-producing democracies. The new organization creates a China-anchored track for a different constituency: countries that have felt underrepresented in Western-led AI governance conversations, along with states that maintain strategic distance from the US alignment. Its membership makes clear a governance track designed for Global South participation from the founding, rather than adding developing-country voices to a framework already shaped by wealthier states.

None of the major Western AI democracies signed. That absence confirms the split reflects strategic division on AI governance, not a wide-participation effort that happens to be China-hosted. Diplomatic and technical coordination on AI now runs through Shanghai as well as through Brussels and Washington, and whichever body moves fastest to specify concrete technical standards (whether that’s model safety testing or cross-border data flows) is well positioned to shape what production AI systems have to look like to operate across borders.

What It Means for Us

Which set of AI rules we live under is starting to depend more and more on where we live. For those of us in the founding member states, our governments have committed to a specific multilateral framework whose norms may diverge from those developed under Western-anchored bodies. For those of us in the countries not signed on, our AI models and platforms will still be exported to and shaped by the standards of the countries where they operate. The AI systems we use every day, from search assistants to the recommendation engines behind our feeds, are built and trained under regulatory environments that are pulling apart.

When the rules governing training data and content moderation diverge across jurisdictions, the same underlying technology gets tuned differently for different markets. That reaches us through what our tools can and cannot say, and through what kinds of automated decisions they are permitted to make in our lives. Any organization operating across borders, from multinational employers to small businesses running ads abroad, now has more governance regimes to reconcile. Even outside the workplace, the online services we depend on cross national borders, and their availability will be shaped by the regulatory maps that took a new shape this week.

Bottom Line: The next question for the new organization is whether it produces binding technical standards or stays advisory. If the new body eventually adopts binding technical standards, that would create structural divergence in how AI gets built — with distinct training and content-moderation requirements encoded into models under each intergovernmental track. Advisory status would make it a diplomatic instrument without meaningful operational teeth. The first technical standard it issues will tell us which track this becomes. AI governance has now entered the domain of geopolitical competition, and the frameworks shaping the AI in our lives will be the ones that move fastest to specify what compliance means in practice.

Researchers Publish a Way to Screen AI for Illegal Content Without Generating Any

Read more

Information overload names the situation. Information overwhelm names the response.

In a working day, it can look like: notifications interrupt our thinking, and tabs multiply across the screen. Thoughts we were following slip past each other. By evening, the morning’s activity can look like a scatter of half-completed tasks.

Most of what drives this scatter sits in the environments we work inside. Phones and feeds aren’t neutral tools. They’re products built by companies whose success depends on our attention, and the people who built them drew on what we know about how the brain works. The end result: a red notification, or a headline built to pull a click, is a cue placed where it will trigger a response before we’ve decided to give one. Willpower can absolutely help us resist these cues, but leaning on it alone is a hard way to win, since it sets one person’s conscious effort against systems tuned by studying billions of users.

What works better is learning to recognize the cues for what they are. Every one of them works by pressing on a specific, predictable feature of human attention, and every feature has a name and a history in the research. Once we can see a countdown timer as manufactured pressure, or a bottomless feed as a design that never lets a session end on its own, we know what’s producing the pull toward the screen. From there we can change our surroundings, which works where sheer resistance in the moment often doesn’t.

Behind each cue is one of the brain’s attention and reward systems, the machinery that decides what we notice and what we chase. Each works fine at an ordinary volume of incoming information and strains as that volume climbs. Our core attentional systems evolved long before digital environments, and have changed little relative to the pace of technological change. Each system shows up as a particular feeling we’ll know from experience, and each one becomes workable once we can spot the design cue behind the feeling.

These attentional systems rarely fire alone. On a scattered afternoon several run at once, which is why information overload lands as one undifferentiated fog and not a set of separate, nameable problems. Taking them one at a time is what turns that fog into something we can act on.

When every buzz and badge grabs our attention

A phone lights up across the room, and attention swings toward it before any decision to look. The same swing happens toward a browser tab the moment an unread count appears. Whatever caught us didn’t have to matter, because the swing comes first and the judgment about whether it mattered comes second.

Why this happens:

A system in the brain decides what deserves our notice before we consciously weigh in. Work in the 2000s helped formalize it as the salience network, anchored in two regions, the anterior insula and the dorsal anterior cingulate cortex, that flag which inputs to bring forward. It evolved to catch change that carried consequences, like motion at the edge of vision that could mean a predator or prey. A red dot or a banner mimics that same sudden change — the kind that once meant danger or opportunity — so the salience network responds to these cues as though they carried the same stakes.

What overload does:

The salience network was built for a world where one or two things a day suddenly demanded notice, with a sound or a movement that might carry high stakes. But a phone and a browser now send it hundreds of manufactured alerts a day. The red notification dot is the cue to read here: red is the color our brains associate with alarm, and the dot uses that alarm color to feel urgent even when nothing urgent waits behind the number. Once we see that the urgency lives in the color, not in the message, the dot loses some of its grip.

What helps:

We can’t out-discipline a reflex, so the leverage is in changing what the salience network has to react to. It helps to turn off notification counts, so no red number sits there manufacturing urgency, and to put the phone fully out of sight during focused work, in a drawer or another room, since face-down on the desk still pulls our attention. The salience network will keep scanning for change, but now it simply finds less of it to catch.

When we can’t keep a thought straight

We’re trying to keep a thought steady while a video plays in one tab and a notification waits at the edge of the screen. The idea turns slippery, each piece sliding away before it can join the one before. The mental workspace where thinking happens has run out of room.

Why this happens:

Working memory is the mental workspace for whatever we’re consciously handling right now, and its capacity is startlingly small. George Miller’s classic 1956 work put the limit at around seven items, and later work by Nelson Cowan revised it downward to about four. John Sweller’s cognitive load theory, developed in the 1980s, built on that limit: when the workspace is exceeded, thinking degrades and confusion sets in. Sweller distinguished two kinds of load: the effort a task takes because of its own difficulty, and the extra effort added by how the task is presented to us. A feed with autoplay video and pop-ups loads us with the second kind, demanding attention for its clutter while giving us little in return.

What overload does:

Working memory evolved for a world where the things in front of us were few and changed slowly. Overload puts far more than a handful before us at once, all of it moving, so the workspace overflows and half-built thoughts scatter. Nothing is wrong with our memory; the load simply crossed a limit that everyone has.

What helps:

The move is to clear the load the task never asked for. Closing the tabs we aren't using and silencing notifications frees the workspace, since every open tab and waiting alert keeps a piece of our attention in the background (and we tend to underestimate how much). Working one thing at a time, finishing or parking a task before the next, lets the workspace carry a single job at once. The room to think returns about as fast as the clutter leaves.

When part of our mind stays on the last task

We’ve switched to something new, and a small mental itch stays behind. Attention is technically on the new task, yet part of it keeps tugging back toward the one we just left. That tug isn’t preference or nostalgia; a piece of our mind is still committed to the last thing, and no act of will releases the commitment.

Why this happens:

Sophie Leroy named this in 2009: attention residue. When we switch tasks, part of our mental capacity stays with the last one, so the next task starts already partly drained. We can’t feel the drain directly, because the capacity still tied up feels available to us. Leroy found that unfinished tasks leave the most residue, the mind circling back to whatever was left open.

What overload does:

The pull back to an unfinished task once helped our ancestors return to things that still needed doing. Modern tools rarely let anything finish: a feed has no end, and a notification breaks one task before another is closed. So the residue never fully clears, and each task runs on a mind already partly committed somewhere else.

What helps:

The residue clears fastest when the mind gets a sense that the unfinished task is handled. Before switching, taking under a minute to jot down where we stopped and the first step for coming back tells the brain the task is safely parked, which is the permission it needs to let go. When the pull lingers anyway, a short gap between tasks helps more than pushing straight into the next thing. The residue fades on its own given a moment of undemanding attention — so a walk to refill a glass of water lets it clear, while opening the next browser tab keeps the mind loaded.

When we can’t stop refreshing

We refresh a feed we refreshed a minute ago, then check an inbox we just checked. We catch ourselves doing this, but the catching changes nothing. Something keeps pulling us back even though we know there’s nothing new.

Why this happens:

The pull comes from dopamine, which many people picture as a pleasure chemical and which works more like a motivation driver. Research on dopamine neurons has found that they fire most strongly in the moment before a reward, when one might arrive and the outcome is still unknown, and that their activity drops once the outcome becomes predictable. Uncertainty is what sets off the strongest dopamine response, and that response is what we feel as the urge to check. B. F. Skinner showed in the 1950s that unpredictable rewards produce the most persistent behavior of any pattern, the principle slot machines run on and now every feed.

What overload does:

At ordinary volume, the dopamine system kept our ancestors seeking food and water when neither was guaranteed. Overload hands it an endless supply of maybe-rewards, so the seeking never reaches a natural stop. Nothing about the system has broken; it’s doing the job it evolved for, inside an environment built to keep it running without end.

What helps:

The urge feeds on uncertainty, on not knowing whether this check will turn up something good, so the way to weaken it is to make each check predictable. Email set to arrive in fixed batches means every check returns the same known result. A timer on scrolling ends the session at a set point, so the feed can't dangle the next maybe-reward indefinitely. This pull is stronger for some of us than others, and for brains with lower baseline dopamine, including ADHD, the environmental change matters most, since willpower has the least to work with there. The goal is a smaller, weaker impulse, even on the days it doesn't disappear.

When our focus is gone by mid-afternoon

Late in the day when work demanded steady focus, a heaviness sits behind the eyes that coffee doesn’t lift. The concentration that came easily at ten in the morning won’t answer at three in the afternoon. The thing worn down is focused attention, and resting the body doesn’t restore that specific kind.

Why this happens:

Attention comes in more than one mode. Rachel and Stephen Kaplan at the University of Michigan separated two kinds: directed attention, which takes effort and fends off distraction, and involuntary attention, which switches on by itself when something gently catches the eye. Directed attention is what most information work runs on, and it tires like a muscle. The Kaplans found it recovers through what they called soft fascination, gentle input that occupies the involuntary kind while the directed kind rests, which natural settings supply well.

What overload does:

Directed attention was built to work in bursts, with recovery arriving through the natural lulls of a day. Overload strips out the lulls, asking for steady focus across hours with no gentle stretches between, so the capacity drains and never tops back up. The wall we hit at three in the afternoon is directed attention pushed well past the recovery it was designed around.

What helps:

What recovers is the exact kind of attention that’s depleted, which sleep by itself won’t cover. When the slump hits, a short walk without the phone, with our eyes settling on something undemanding like trees or open sky, gives directed attention room to recover while the involuntary kind idles. These low-effort stretches are where the restoration happens. A break like this guards the concentration the rest of the day leans on.

When we look up and twenty tabs are open

We opened one link an hour ago. Now twenty tabs are open, none of them the thing we came for, and the question that started the search has vanished into the branching. The instinct behind this is old, the pull to follow a promising trail wherever it leads, and in these surroundings the trails never run out.

Why this happens:

Looking for information behaves a lot like foraging for food. Peter Pirolli and Stuart Card described this in the 1990s, borrowing from studies of how animals find sustenance, and called the cues we follow information scent. We read a link’s wording and guess whether the path will pay off, following the strong cues and dropping the weak. The instinct is efficient whenever the cues are reliable.

What overload does:

Foraging works when a strong cue means a rich find. Overload floods the surroundings with headlines built to “smell” richer than they deliver, so cue strength stops matching payoff; with endless paths on offer, the foraging spirals into a screen full of tabs. The headline is the cue to read here: strong scent is now something a platform can manufacture, so a link's pull no longer tells us much about what actually waits at the end.

What helps:

Foraging has no natural stopping point online, so we have to bring one: a specific target, fixed before we start. Naming exactly what would count as the answer gives the search an endpoint; once that endpoint exists, the tempting side-paths show themselves as side-paths we can pass by. Stopping the moment we reach the target, even with links still beckoning, keeps the search from branching. A search with a finish line ends; a search without one wanders.

When our best ideas arrive in the shower

The solution we couldn’t reach at our desk arrives in the shower. The name we couldn’t summon surfaces on the walk to the kitchen. Both are traces of a mode of thinking that most of the day leaves no opening for.

Why this happens:

The brain has a mode that switches on when nothing outside is asking for attention. In the early 2000s, brain imaging research identified a pattern of areas that becomes active when people are at rest and named it the default mode network. It's associated with mind-wandering, and is thought to support work like tying recent experience to what we already know and forming the connections that later surface as insight. In this mode, when it gets the chance to run, the day itself gets processed.

What overload does:

Mind-wandering evolved as the stretch where experience gets integrated and loose connections form. Overload fills every gap with input, so the default mode network rarely gets its opening, and the integration never begins. Some of the missing insights and forgotten details may come from exactly this: a mode that keeps getting crowded out before it can do its work.

What helps:

What this mode needs is stretches of empty time, with nothing scheduled and nothing playing. Showering without a podcast, or taking the occasional walk without headphones, gives the mind room to wander. The ideas that rise in these gaps are the product of the gap itself. A little guarded emptiness feeds insight in a way that extra input never will.

Seeing the systems for what they are

Everyone with a phone and a feed lives inside these same systems. A scattered afternoon, or a three o’clock collapse of focus, reflects the environment we spent the day in; these systems are standard equipment in every brain, and the platforms pressing on them work the same way for all of us. A day that got away from us was never a measure of our discipline.

A pattern we can name is a pattern we can work with. When refreshing shows itself as a reward system chasing an uncertain payoff, we can hand it a predictable schedule and watch it settle. When focus gives out under sustained effort, a short walk outside restores it in a way another coffee at the desk can’t. Each fix meets one system on its own terms, and the relief adds up across a day.

Information overload isn’t a personal failing to be willed away, and the overwhelm that comes with it isn’t a sign we’re weak. Both are part of the world we live in now, a world deliberately arranged around capturing our attention. Learning to see how that works is bigger than any single fix for a hard day. It’s a way of moving through the world with our eyes open to how an environment is designed and to what it wants from us. That same clear sight applies well past our phones, to anywhere built to steer what we do next. Overload is the condition of the world we’ve been handed. But what we get back by seeing it clearly is room to think: the attention to follow an idea to its end, and to have thoughts that are ours and not just reactions to whatever reached us first. That clarity is what lets us live in this world on our own terms.

You might also like:

• We Were Never Supposed to Know This Much: On the cost of knowing everything, everywhere, all at once.

• From Ataris to AI: The Millennial Reckoning With Relentless Change: What it means to grow up analog, come of age digital, and arrive at something no one has a word for yet.

• What Scrolling Did to Reading: Why books feel harder, what the science shows, and how to rebuild the capacity.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides. Upgrade to paid for the full Card Catalog.

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

Government reports on the divide started arriving in 1995, when the U.S. National Telecommunications and Information Administration released “Falling Through the Net,” showing how income, geography, race, and education carved stark boundaries around technology access. President Bill Clinton stood with Vice President Al Gore in 1996 and pledged to connect every classroom to the internet by the turn of the millennium under the “Building a Bridge to the 21st Century” banner. The European Commission launched its eEurope initiative in December 1999, aiming to build an “information society for all.” South Korea announced its Cyber Korea 21 strategy that same year, committing to bring citizens online regardless of age, gender, region, or income. Each initiative treated connecting more people to the internet as foundational to broader access.

These bridges got built on a single assumption: that once people were online, the hard part was over and that broader access would inevitably follow. But the pandemic showed what that belief had missed: when schools across nearly every country moved online almost overnight, hundreds of millions of children with no computer or home internet were shut out. Teachers spoke into grids of darkened squares while parents shared cracked phones and stretched thin data, doing what they could to keep them learning on whatever connection they could find. The pandemic didn’t create these gaps; it only made them impossible to ignore.

The gap this exposed was an old one. Underneath every new technology sits the same question: who can navigate the information system in front of them, and who can’t?

The four layers of digital inequality

That question breaks into four layers, each a different way an information system can include or exclude us. The idea comes from two decades of research. The Dutch sociologist Jan van Dijk mapped access to digital technology as four successive modes: motivation, physical means, skills, and use. Work by Eszter Hargittai and Paul DiMaggio pushed the field to study what people can do once they’re online, since skill gaps persist long after a connection is in place. The four layers below build on that scholarship, and they apply to a dense government form or a bureaucratic office as much as to the internet. Recognizing them helps us see what issues stay the same as the technology keeps changing.

Physical access is whether someone can get connected at all. Internet use reaches about 94% of the population in high‑income countries versus around 23% in low‑income ones, and the urban‑rural divide remains particularly stubborn in low‑income regions. Similar patterns reproduce themselves at smaller scale within wealthier countries. A 2024 ITU analysis notes that more than half of the world’s population now lives in areas covered by 5G but warns that universal connectivity remains a “distant prospect,” especially in low‑income economies. In the United States, a February 2026 analysis by the broadband nonprofit Connected Nation estimated that nearly 16 million Americans still lacked access to broadband infrastructure, while another 15 million lacked any internet‑connected device at home. For a student living blocks from a wired campus, that can mean doing homework in a parking lot, borrowing wifi meant for store customers because no connection ever reached their home.

Skills and literacy extend far beyond basic computer operation. Using the internet safely means recognizing scam websites before handing over credit card numbers, and knowing how to adjust privacy settings to limit the data that companies harvest. These skills are a moving target, since each new technology assumes fluency with the last, and the gap becomes apparent the moment a software update or an unfamiliar interface makes hard‑won habits feel useless. Someone who learned computers in the 1990s might handle email with ease yet stall when a workplace moves every file to cloud storage locked behind a two‑factor login and collaboration tools they’ve never used. This gap falls unevenly. Around the world, about 77% of men use the internet compared with about 71% of women, and the divide is widest in low‑income countries. Disability and language cut deeper still, since screen readers and translation features remain afterthoughts in much of what gets built, especially for languages with fewer speakers or non‑Latin scripts.

Quality of access determines what becomes possible after connection exists. A teenager doing calculus homework on a phone with a cracked screen and a spotty data plan inhabits a different technological universe than a peer with a laptop and dedicated home wifi. Connection speed affects which applications run smoothly, while stability governs situations such as when a video interview can happen without dropped calls. 5G coverage remains heavily concentrated in high‑income countries, and many rural areas in the Global South still depend on older networks where video calls drop and webpages load slowly.

Meaningful use is the divide that persists even after access barriers come down. Two people with similar devices and connection speeds can end up in radically different places based on what they do with those tools. One person might learn to code through free online courses or build a professional network that opens unexpected doors, while another uses the same infrastructure mainly for entertainment. Meaningful use depends on the invisible curriculum taught by the people around us, which means the divide tends to perpetuate itself through social learning when people without exposure to certain uses can’t easily imagine those uses exist.

Why the gap keeps growing

The income gap reveals how these layers compound at every scale. ITU’s 2024 report notes that in low‑income countries, the cost of a fixed broadband subscription still represents a large share of average monthly income, while the same service is a smaller fraction of income in wealthier economies. In the United States, Pew Research Center data from 2025 show that only about 54% of adults in households earning under $30,000 subscribe to home broadband, compared with around 94% of adults in households earning over $100,000. Even in the world's wealthiest countries, that gap shapes who can find work and reach care as those services move online.

Recruitment shows the consequences of that gap directly. Many industries have shifted most hiring to digital platforms, with large employers often refusing paper submissions. Companies screen applicants through automated systems that parse resumes for keywords before human eyes ever see them, and a welder with twenty years of experience can watch her application disappear because the software didn’t find the right phrases. Remote work opportunities remain inaccessible to people without home internet and appropriate devices, which turns connectivity into a major factor in expanding the geography of opportunity.

Healthcare has transformed just as dramatically. A patient who once drove ninety minutes to see a specialist might now have the appointment at home if the connection cooperates. Data from the U.S. National Center for Health Statistics show that the share of office‑based physicians using telemedicine technology rose from about 16% in 2019 to nearly 86% in 2021. But patients without reliable internet access can’t easily tap into telehealth services, forfeiting convenient care and the ability to consult specialists their region doesn’t have.

Government services have followed the same trajectory online. India's Aadhaar system links welfare payments to bank accounts for more than 1.4 billion people, and comparable programs have spread from Estonia to Brazil. Each one serves the people who can navigate it, and leaves those who can’t behind. And in the United States, the Affordable Connectivity Program, which had subsidized broadband for around 23 million households, ran out of funding in 2024 - just as more of these services were moving online.

How AI amplifies existing inequalities

Into this widening gap, artificial intelligence has arrived with a velocity that few previous technology transitions match. AI amplifies all four layers of the divide (access, skills, quality, and meaningful use) at once. Earlier waves required access and basic literacy, with skills that remained relatively stable once acquired. But AI tools shatter that pattern. Their capabilities expand every few months, which means the literacy required to use them well never settles. By the mid‑2020s, the landscape includes rapidly evolving models from OpenAI, Anthropic, Google, Mistral, DeepSeek, and others, with new releases arriving fast enough to make any learnings quickly outdated.

Language adds a layer of its own. The leading AI models learn mostly from English. In a language with less text online, like Swahili or Bengali, the same question tends to return a thinner, less accurate answer. Labs in France, China, India, and elsewhere are now training models on their own languages, which means who builds an AI increasingly shapes whom it serves well.

AI tools also require different cognitive skills than previous technologies, building atop existing digital literacy while demanding higher‑order thinking. Using a search engine effectively means knowing how to formulate queries and evaluate results, whereas using a language model effectively requires those same skills plus the ability to articulate complex requests clearly and to check outputs for accuracy and bias. The evaluation skills that help us decide whether a website is trustworthy carry even more weight with AI‑generated information, where systems can fabricate confident‑sounding answers entirely.

Even schools are splitting the same way. Wealthier U.S. districts moved fast to train teachers on AI in 2023 and 2024, while poorer ones fell behind. RAND projected that nearly every affluent district would have done so by the fall of 2025, compared with about six in ten of the poorest. That leaves students in poor districts with less help learning the tools everyone else is starting to use.

Cost raises the barrier further. The most capable AI tools sit behind monthly subscriptions that run from about $20 for a basic plan to $200 for the top tier - more than a month's income across much of the world. Free versions exist, but they come with catches. These run on older, weaker models with tight limits on use, and they often feed what people type into training (so that access is paid for in data). That difference also compounds, since faster and sharper answers help someone work and learn more, which in turn pays for the next upgrade. What begins as a price gap hardens into a gap in what the tools can do, and the people who stand to gain the most from a strong assistant are the least able to reach one.

The sharpest concerns appear where AI already makes decisions about people. Automated hiring systems filter job applicants using algorithms whose training data may encode decades of discrimination, while credit scoring algorithms make lending decisions that can perpetuate racial and economic inequality. Facial recognition systems have been shown to work less reliably on darker skin tones because they were trained predominantly on lighter‑skinned faces, with documented harms including misidentification and wrongful arrests. Each of these failures tends to hit the same populations the digital divide already disadvantages.

Governments have started to regulate AI and expand connectivity in very different ways. Some, like the European Union, have written detailed rules about how AI can be used, while others have focused on building infrastructure or left the harder questions mostly unanswered. Those choices shape whether schools, clinics, libraries, and other public institutions can use AI and broadband to narrow the divide, or whether they’re left to fill the gaps that policy never reached on their own.

What communities can do

Filling those gaps often starts with the institutions closest to a community. These groups know which neighborhoods lack service and where residents need help getting online, detail that national programs rarely capture. Library boards, school boards, community organizations, faith communities, and local councils can each narrow the divide within their reach.

Expand library and community center programming.

Public libraries already serve as community technology hubs in many places. Expanding this role with dedicated AI‑literate staff and on‑site access to premium AI subscriptions could help millions develop new skills, especially when paired with evening and weekend programming that meets working adults where they are.

Build AI literacy into local schools.

Students benefit from hands‑on experience using AI tools while developing a sophisticated understanding of ethical use boundaries under the right guidance. The skills include fact‑checking AI outputs against primary sources and combining AI help with human judgment rather than outsourcing their thinking entirely to algorithms. Training data shapes AI behavior in ways that encode historical biases, a pattern students learn to spot. School boards and parent organizations can press for curricula that build these skills alongside reading and mathematics.

Fund community‑led solutions.

Detroit’s pandemic response mobilized schools and community organizations to distribute hotspots and devices in ways that fit local needs. Kenya’s mobile money ecosystem grew from a mid‑2000s pilot into a system that reshaped financial inclusion across an entire region. Locally responsive work tends to outperform top‑down programs designed without input from the people they aim to help, and community organizations already doing this work are where funding goes furthest. Mutual aid networks, community organizations, faith communities, and disability advocacy groups have built trust and cultural competence over years that new programs can’t quickly replicate.

What governments can do

Some of this work sits beyond the reach of any individual or community, though, in the domain of national policy and public spending. Governments can act at a scale no one else can, funding broadband across whole regions and setting the rules for how AI is applied in hiring and lending. Those are the decisions that will shape whether the divide narrows or widens.

Build infrastructure to reach everyone.

The gap in physical access sits mostly in rural and low-income areas that commercial providers pass over. Municipal networks like Chattanooga's show that community‑owned broadband can match or beat what private companies deliver, often in exactly those places. Affordability is the other half of the job, and it is where subsidies like the Affordable Connectivity Program come in.

Provide public AI as civic infrastructure.

The civic principle that lets anyone walk into a library and use a computer extends to AI tools as well. Government agencies at every level can develop AI assistants that guide people through processes like applying for benefits in accessible language and in the languages communities speak. With funding, adult education programs can offer AI literacy courses to working people who can’t return to school full‑time.

Mandate transparency and accountability for AI that shapes people’s lives.

Countries diverge sharply on AI, from the European Union’s single law that puts the toughest limits on the highest-risk uses to the United Kingdom’s lighter, industry-by-industry approach. Rules can require companies to reveal when an automated system helps decide someone’s job, loan, housing, medical care, or school placement, and to give people a way to challenge the outcome. Checking a tool for bias before it launches, along with review by outside auditors, can catch problems a company might otherwise miss or minimize.

What the bridge was always for

Every information system invents its own architecture of access, from printed catalogs to microfilm and search engines to language models. Each format has produced gaps between people who can navigate the system and people who can’t. The internet now reaches about three‑quarters of humanity, and the next technology is likely to rebuild that gap underneath us but in a new form. Every new bridge resurfaces the same question those late‑1990s reports raised: who gets left behind?

AI is raising the next bridge now, at the same early stage the internet occupied in 1995. None of its layers has hardened yet. But unlike 1995, the map of the divide already exists, with three decades of research showing where gaps form and how they deepen. This time the people working on the problem know what to watch: the cost of the tools, the skills to use them, the quality of the connection, and what people do once those pieces are in place. All of that amounts to a head start: a chance to narrow the gap while it’s still forming.

You might also like:

• How to Teach Kids to Evaluate Information (Before AI Teaches Them Not To): An information literacy primer for parents and educators.

• The information diet isn’t enough. What we need is an information village.: What changes when we go beyond optimizing our consumption and start building shared infrastructure.

• How to Use AI to Make Us Smarter: The headlines say AI is making us dumber. The research tells a more complicated story.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Briefings, monthly Librarian Hotlines, and quarterly deep-dive guides. Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

This week:

• Until now, catching an AI system in a dangerous mistake meant having nowhere to turn. A group of researchers just changed that, and Congress is taking note.

• Most of the art world has turned against AI, convinced it means the end of human creativity. Two of its most influential figures are staging the opposite case: that the technology could touch off a renaissance.

• The U.S. and China are racing to build the most powerful AI. But China is also fighting a second battle the U.S. has mostly ignored: keeping the technology from throwing its own people out of work.

Researchers Launch a Shared System for Reporting AI Flaws

What Happened

A group of AI researchers launched FLARE-AI, a free website where anyone can report a flaw or harm they run into while using an AI system. The name stands for Flaw Reporting for AI, and it works as a single place to raise the alarm when a chatbot does something it shouldn’t (from handing a user malware to steering a vulnerable person deeper into delusion). Until now, someone who spotted a serious problem in a chatbot had no obvious place to send it and no assurance the right people would ever find out.

The project was led by Avijit Ghosh, a policy researcher at the AI company HuggingFace, with computer scientists Elaine Zhu and Shayne Longpre, and it was shaped by 49 experts from 32 organizations. Because the site’s code is open for anyone to inspect, an independent reviewer can confirm a reported problem and route it to the company that built the model, along with organizations like MITRE that catalog failures in technical systems. The result works a bit like Downdetector, the service that aggregates user reports when an app or website goes down. The same team also advised on a bill introduced in Congress in June that would put a US government agency, the National Institute of Standards and Technology, in charge of writing reporting standards and maintaining a central database of AI flaws.

Why It Matters

For as long as modern AI has been in public hands, the companies building the systems have also been the ones who decide what counts as a flaw and how much of it the public ever hears about. A researcher who found a serious problem could report it to the company and hope, but there was no shared standard for what a report should even contain, and no way for anyone outside the company to confirm the problem or carry it to others facing the same risk. What FLARE introduces, and what the congressional bill would harden into law, is the opposite arrangement: a common format for reports and a route that carries them past any single company’s discretion.

This is the same shift computer security made a generation ago. Software vulnerabilities were once treated as private embarrassments, patched at the vendor’s discretion or hidden outright, until the field settled on coordinated disclosure, a public and standardized system that treats a security hole as shared safety information instead of something a company keeps to itself. AI has had no equivalent, even as these systems moved out of the lab and into places where their mistakes carry consequences.

What It Means for Us

We rely on these systems now without any real way to check them. We ask a chatbot for medical advice or hand it a document full of personal detail, and we trust that the company behind it has already caught whatever might go wrong. That trust has been the only option, because the inner workings are closed to us and the record of failures has lived inside the companies themselves.

A shared, independent reporting system starts to change our footing. The flaws that could do the most damage are the ones that are difficult to catch ourselves. For example, a model that can be tricked into handing over private data, or one that steadily feeds a person’s darkest thoughts, is exactly the kind of failure this system exists to catch and pass to someone who can fix the problem. Now that someone outside the company is finally watching, the tools we already rely on get safer as accountability increases.

Bottom Line: Reporting a flaw does not oblige a company to repair the flaw. The tools we depend on grow safer only when a company has a reason to act on what gets reported - and a website alone cannot create that reason. FLARE gives us a way to surface problems, while the power to make a company address them stays out of reach. The congressional bill is where that could change, because a public database of flaws a company was warned about and never fixed would let us see which companies leave problems unaddressed. Once that record exists, a company's safety claims become something we can check rather than something we accept on faith, and we can move our trust toward the ones that follow through.

Two Artists Make the Case That AI Can Renew Art

Read more

Watching this dynamic unfold around him, Jonathan Swift made an observation that would outlive every pamphlet of his era. In a line widely attributed to him, Swift wrote: “Falsehood flies, and the Truth comes limping after it; so that when men come to be undeceived, it is too late; the jest is over, and the tale has had its effect.” Swift was describing something more enduring than any particular medium. He named the asymmetry between the speed at which a false claim can travel and the speed at which careful evaluation works. That asymmetry shaped public life in 1710, and it shapes ours now in ways that would be unrecognizable to him but operate on the same principle.

Three hundred years later, a team of MIT researchers put numbers to what Swift had observed. Soroush Vosoughi, along with Deb Roy and Sinan Aral, analyzed roughly 126,000 stories shared on Twitter over more than a decade. False news reached large audiences significantly faster than accurate news, on the order of several times faster to reach 1,500 people. Falsehoods were 70 percent more likely to be retweeted than the truth, and the effect held across multiple categories of information the researchers tested - from political news to natural disasters to urban legends. The team also ran the analysis with all the automated accounts stripped out, expecting the gap to shrink. The gap barely budged at all, which meant the spread was being driven by human users rather than by bots.

The researchers suggested that novelty helps explain the pattern. False news, they found, was more novel than accurate news. We respond to novelty with surprise and sharpened attention, emotions that heighten engagement and are readily shared with others. Sharing novel information confers a kind of social currency: it makes us look informed and plugged into the moment. The incentive structure of a social platform often rewards early sharing and rapid engagement, and false claims, unencumbered by the friction of verification, tend to move faster than true ones.

When we learn that someone we know has shared something that later turned out to be false, the instinct is to look for the flaw in their thinking. We assume they lacked the critical thinking skills to see through the claim. But the MIT data points in a different direction. The speed of the encounter, shaped by the design of the platform, appears to play a decisive role in whether our critical thinking gets used at all. The same person who would catch a false claim when reading carefully can miss it entirely when scrolling fast. What changed isn’t the person but the pace.

Why we share false headlines

The MIT data established what happens when false claims meet the speed of a social feed. What it didn’t fully explain was why any given one of us falls for a false claim in the moment of encounter. A few years before the MIT study, two cognitive psychologists set out to test the dominant theory of the time, which held that people believe false claims because the claims confirm what they already want to be true, and because partisan loyalty overrides their capacity for careful evaluation. Gordon Pennycook and David Rand designed studies to find out whether that theory held up.

Pennycook and Rand found that susceptibility to false headlines was strongly related to whether people engaged in analytical thinking at the moment of encounter, and less strongly explained by partisan alignment than many theories had assumed. The key factor was not whether people were capable of reasoning well, but whether they engaged that reasoning in the moment. Participants in their studies were broadly capable of distinguishing false claims from true ones when the conditions prompted them to try, but absent those conditions, the analytical capacity sat unused while faster, automatic judgments handled the evaluation instead. Their 2019 paper in the journal Cognition captured the finding in a memorable subtitle: “Lazy, not biased.”

The word “lazy” sounds harsher than what the researchers mean by it. They use it to describe a predictable cognitive response to an environment engineered for speed, not a moral judgment or a lack of effort. When we move through information at the pace a social media feed sets for us, the deliberate thinking that would help us distinguish a reliable claim from a fabricated one doesn’t engage. The capacity for careful evaluation remains present, but nothing in the environment prompts its activation. A key variable that shapes whether we catch a false claim turns out to be the conditions of the encounter, not the cognitive equipment we bring to the task.

For decades, the dominant response to misinformation has been educational. If we could teach people to evaluate sources more carefully, the thinking went, they’d be harder to manipulate. But the Pennycook and Rand research suggests this approach may be incomplete. Their findings indicate that the speed of our information environments can prevent critical thinking from activating in the first place. The intervention point isn’t more education in how to think, but more conditions in which our existing thinking has a chance to engage at all.

How the feed outpaces deliberation

The mechanism is rooted in how our minds handle two different kinds of thinking. Psychologist Daniel Kahneman, in his book Thinking, Fast and Slow, described cognition as if it operated through two modes. The fast mode is automatic and pattern-based, running on emotional responsiveness and prior experience. It does the everyday work of recognizing a friend’s face from across a room or registering the gist of a headline in a single glance. The slow mode is deliberate and effortful, engaging when we pause to weigh a claim against available evidence or notice that something doesn’t quite add up. The two modes work in tandem, with the fast system generating impressions that the slow system can accept or override. But the slow system depends on something the fast system doesn’t: a pause in the encounter.

Social media feeds are designed in ways that keep the fast system running without interruption. The infinite scroll, the default interface for nearly every major platform, removes the natural stopping points that once shaped how we consumed information. The bottom of a printed page or the end of a broadcast used to create involuntary pauses in our consumption. Without those structural breaks, consumption becomes continuous, which grants the fast processing system an unchecked advantage. Algorithms often prioritize emotionally engaging content, which tends to increase user engagement and, in turn, platform revenue. Outrage and delight both work as fuel that keeps the feed moving and the user inside the app. And every feature that reduces the friction between encountering a claim and broadcasting it, from the one-click retweet to the frictionless share button, compresses the window in which the slow system could intervene.

These platforms were optimized for engagement, and the effect of that optimization is an environment where the deliberation that would help us evaluate a claim before spreading it rarely has a chance to engage. The faster we move through content, the more content we engage with, and the more data the platform collects about our preferences and behaviors. Every communication technology since Swift’s era has accelerated the spread of false claims, but the human work of evaluating a claim still takes the same effort it always has. Radio compressed the distance between speaker and audience, letting one voice reach millions of listeners at the same instant. But radio still kept that audience on the receiving end of the broadcast. Social media collapsed that arrangement entirely, taking the speed of broadcast and handing it to every user, so that any individual encountering a claim could amplify it to hundreds of strangers with a single tap.

The power of a single pause

The “Lazy, not biased” finding pointed toward an obvious experimental question: if the problem is that analytical thinking fails to engage at the moment of encounter, would a small intervention designed to engage it change behavior? Building on that insight, Pennycook and Rand and their collaborators ran a series of experiments to test what they called an accuracy prompt. Before participants engaged in a simulated social media sharing task, the researchers showed them a single unrelated headline and asked them to rate whether it seemed accurate. Whether participants rated that headline correctly was beside the point of the study, since the intervention worked by getting people to pause and consider accuracy at all. The act of holding attention on the question for one headline, even a headline unconnected to anything else in the session, was the core of the intervention.

Across 20 experiments involving more than 26,000 participants, the accuracy prompt reduced false news sharing by roughly 10 percent on average. The effect held across political affiliations and across content categories. The structured pause asked the slow system to engage before the fast system took over for the rest of the session. The researchers named the underlying phenomenon inattention, by which they meant participants could distinguish true from false when the environment prompted them to try, while the feed itself never offered such a prompt of its own accord.

A roughly 10 percent reduction in false news sharing sounds modest until we consider what produced the change. The intervention was extremely low-cost and easy to scale, and it required no behavioral change beyond a single moment of attention. Some replication studies have produced smaller effect sizes than the originals, but the basic finding has held across multiple research teams and content types. If a single prompt at the start of a session can shift the quality of every sharing decision that follows, the same principle could shape how we engage with information on our own.

Patience as information literacy

The accuracy prompt works in a controlled setting because researchers can engineer the pause. Outside the lab, we have to build it for ourselves. That work, repeated across our daily encounters with information, is the practical form of information literacy in the age of the feed. When we talk about information literacy, we tend to focus on skills like evaluating sources and cross‑referencing claims. The Pennycook and Rand research suggests these skills only activate under certain conditions, and speed appears to be one of the key conditions that can shut them down. A thorough understanding of how to evaluate a source does us little good if the speed of the encounter prevents us from applying that understanding in the first place. Information literacy has always been about more than the skills themselves; it encompasses the structures that organize information and the pace at which we encounter it, both of which shape what we can do with what we find.

Patience then becomes an information literacy practice in its own right, something we can cultivate and strengthen through repetition. Practicing patience with information involves learning to interrupt one habitual behavior: the automatic forward motion of the scroll. The pause itself is the practice, and the practice works on a single decision point we encounter many times a day.

Three practices emerge directly from the research:

Open the article before sharing when a headline triggers a strong emotional response.

The feeling of surprise or outrage is often a sign that a claim is novel and provocative, which is also what tends to make false claims travel fast online. If a headline makes us want to share it immediately, that urgency is itself a cue to pause. Opening the article and reading past the headline is often enough to change our assessment of whether we should amplify the claim.

Check the claim from a second source before sharing when a headline matches what we already think is true.

A claim that agrees with our existing view of the world arrives with a built‑in feeling of credibility. We scrutinize it less and pass it along faster, because the work of evaluation feels unnecessary. That feeling is exactly what false claims tailored to a particular audience are designed to produce. Looking up the same claim from an unconnected source is one of the most reliable ways to find out whether the claim has evidence behind it, or whether we simply agreed with it because it fit our expectations.

Read carefully even when we don’t plan to share, for our own understanding.

A claim doesn’t need to be passed along to change what we believe. Headlines we scroll past and articles we skim still contribute to the picture of the world we carry into every encounter that follows. The slow‑reading practice applies whether we share or not, because the encounter itself is doing work on us either way. The claims that affect us most are the ones we owe the most careful reading, regardless of whether anyone else ever sees us engage with them.

The more we practice these, the more habitual they become. What starts as a conscious interruption of the feed’s pace turns, over time, into a different relationship with information altogether. The pause stops feeling like an extra step layered onto reading and becomes part of how we read in the first place. We begin to notice things we used to miss. An unfamiliar byline starts to register where it used to slip past. A clipped quote begins to feel suspicious, and the gap between what a study found and how a headline characterizes its findings becomes visible to us. Catching things like these is what attention does when it has the time and space to do its work.

What slowing down gives us back

What returns to us when we slow down, more than anything the research can measure, is our own judgment. The capacity for careful evaluation has been with us the whole time. The speed of the feed has been the variable keeping that capacity from doing its work, and a slower encounter restores the conditions in which the work can happen. Our understanding of the world begins to be shaped by our own thinking about what we’ve read, rather than by the residue of impressions that fast scrolling leaves behind.

Information literacy has always been about being able to think for ourselves about what we read, rather than absorbing whatever happens to cross our path. The point of evaluating sources and cross‑referencing claims is to keep us in that thinking position. Those methods presuppose a reader whose judgment is already engaged when the reading begins. The slowing‑down that returns judgment to us is what makes that larger aim reachable. We become readers present at the encounter, with our attention engaged, holding the power to decide what we take on and what we set aside.

Every time we slow down to engage that way, we also close, by a little, the gap Swift wrote about in 1710. The dynamic he described has shaped information for three centuries since. He could not have known that the gap between false flight and limping truth would one day be measured in milliseconds of scroll, or that the platforms carrying most of our information would be designed in ways that keep the gap from closing. The closing of the gap will not come from the platforms or from the systems built around them. It will come from us, in the moments between one claim and the next, every time we choose to slow down and think before reading on.

You might also like:

• What Scrolling Did to Reading: Why books feel harder, what the science shows, and how to rebuild the capacity.

• 20 Small Acts of Media Literacy We Can Practice Every Day: From checking the date to tracing the source: building a sharper eye for information.

• The Hierarchy of Sources: A Cheat Sheet: A guide to evaluating information sources in the AI age.

The free essays are the foundation. The paid tier is where information literacy becomes a skill you can use: the Card Catalog Classroom with its searchable Reference Collection and structured lessons across four learning domains, plus biweekly AI Intelligence Briefings, monthly subscriber-driven research, and quarterly deep-dive guides. Upgrade to paid for the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

Getting there isn’t simply a matter of asking better questions. It takes a tool that’s willing to engage with our thinking, and by default, ChatGPT and Claude aren’t built for that kind of engagement. Both are designed to be agreeable. They affirm shaky reasoning and generate polished answers to half-formed questions, rarely pushing back even when pushing back is what would actually help us. That agreeableness is a structural obstacle to using AI as a thinking partner, and it shows up in every conversation we have until we tell the AI to behave differently.

That’s where custom instructions come in. They’re the first place we tell the AI what kind of working relationship we want, and the most direct way to shift the dynamic from agreeable assistant to actual collaborator.

Reader Question

“I use ChatGPT a lot for work, and lately I’ve started to wonder if it’s making me worse at my own job. I’m not working through problems the way I used to, and I’m taking everything it tells me at face value. Is there a way to set it up so I’m still involved in the work instead of just outsourcing it?”

— J.R.

A note before we begin: Thank you for being here, whether you’re a free or paid subscriber. This post covers what custom instructions are, where to find them, and the principles that determine whether the AI keeps us mentally engaged or lets us coast (available to everyone). The step-by-step setup in both Claude and ChatGPT, a complete starter set of instructions to paste into either platform, and how to extend the same thinking-partner setup into Projects for ongoing work are available for paid subscribers.

Librarian Answer

The concern in this question is one of the defining themes of our current moment. AI tools are powerful enough to handle the analytical work we used to do ourselves, and using them constantly without the right guardrails can leave our own reasoning a little weaker each time we open a chat. The reflex to just accept what comes back can be hard to resist, but the cost of that reflex compounds. Once we stop asking our own questions, the AI becomes the only voice in the conversation.

The way out is to change how the AI shows up in the conversation in the first place, and the lever for that is a feature called custom instructions. Both ChatGPT and Claude have a text field in their settings where we can write a short note the AI reads before every conversation begins. What goes in that field shapes the entire dynamic that follows. Used well, it can be the difference between an AI that hands us a finished answer every time and one that asks what we’ve already considered before responding, then pushes back when our reasoning has a gap.

The most common approach to filling this space is with practical details like a job title or a formatting preference. But the same space can go further, changing how the AI responds when we bring it a half-formed idea or a question we haven’t fully worked through yet.

What changes when we use this space differently

Whatever we paste into the custom instructions field becomes a standing directive for the AI, applied to every new conversation. The wording we choose shapes the starting point of those conversations, and what that starting point looks like is up to us. Asking the AI for shorter, faster answers will tend to produce shorter, faster answers. Asking it to slow down and surface gaps in our reasoning will steer the conversations in that direction instead.

Here’s the kind of note someone might write if they want quick, usable output:

“I’m a marketing manager. Give me concise answers.”

And here’s an alternative someone might write if they want the AI to slow them down and engage with their reasoning:

“Before giving me a solution, ask me what I’ve already considered. If my thinking has a gap, point it out before moving forward.”

Both are useful, but they create very different experiences. A direct prompt for concise answers gets us to a finished result faster, which makes sense when we know exactly what we need. A prompt asking the AI to question our reasoning and point out gaps before answering creates a back-and-forth where it slows down and pushes back when something looks off. That kind of back-and-forth is what keeps us mentally engaged in the work instead of waiting on the AI to finish it for us, and it starts with writing instructions that build friction into the process on purpose.

The most useful kind of friction is making the AI ask us something before it gives us an answer. AI can produce a polished response before we’ve had a chance to work through the question ourselves. A polished answer is easy to accept at face value, especially when it’s well-written and addresses what we asked. An instruction like "when I bring you a problem, ask me what I've already tried or considered before jumping to a solution" changes the order of the conversation. Instead of the AI going straight to an answer, it has to check in with us first. That moment, where the question turns back to us, gives our own thought process and judgment a chance to surface before the AI fills the space.

That principle extends naturally to the assumptions embedded in our questions. Without explicit permission to push back, AI will build on whatever premise we hand it, even a shaky one. If we walk in with a flawed assumption baked into our question, the AI will treat it as settled and build from there. An instruction like “if my question contains an assumption that might be wrong, flag it” gives the AI a role beyond answering: checking whether the question itself holds up. It’s a challenge to give ourselves that kind of scrutiny unprompted, which is exactly why it helps to build it into the tool.

Both of these ideas can live side by side in a single short paragraph in the custom instructions field. The result is an AI that stops producing answers on autopilot and starts drawing us into the process, where our own thinking has a chance to shape where the conversation goes.

Setting it up

Read more

The Classroom is a private learning space on Circle where the ideas and resources shared across Card Catalog articles and Notes are organized into a searchable library and built into structured lessons with hands-on exercises. It’s a place to work through skills and explore organized resources at our own pace, with everything available to come back to whenever we want.

The Classroom is a living library, and like any library, the collection grows over time. Each Lesson is built from one of Card Catalog’s articles: where the article explains an idea, the Lesson is a way to put that idea into practice.

New lessons are added regularly, organized across four learning domains:

• AI Literacy

• Source Evaluation

• Research Skills

• Digital Culture

The first lesson available is Use AI to Sharpen Our Thinking, built from the article How to Use AI to Make Us Smarter.

Alongside the Lessons is the Reference Collection. If we’ve ever read a Note, bookmarked a resource, or thought “I know she shared something about this!” and couldn’t track it down again, all of that now lives inside the Classroom too. Every archive, database, tool, framework, and recommendation ever shared in Card Catalog Notes has been organized into a searchable directory of 16 categories:

• AI Prompting Playbook

• Archives and Databases

• Evaluation Frameworks

• How AI Works

• How Algorithms Work

• How We Trust Information

• Information and Library History

• Information Literacy Tip of the Day

• Information Overload

• Librarian’s Dictionary

• Libraries of the World

• Library Science Tip of the Day

• Reading and Discovery

• Research Skills

• Verification Toolkit

• Virtual Museum Tours

Good information is findable information, and the Reference Collection is what that looks like in practice. New resources are added every week as new material is shared in Notes.

We also have an Archives section that holds every AI Briefing, Librarian Hotline, and Research Packet ever published, gathered in one place inside the Classroom. New ones will still go out on Substack as they’re written, and they’ll show up in the Classroom Archives at the same time.

A note for paid subscribers

If you’re already a paid subscriber, you already have access. Circle will be sending an invitation email in the next few days to set up an account. The email comes from Circle directly, so keep an eye on the inbox (and the spam folder, in case it lands there). Once the link is clicked and a password is set, your Classroom is open.

On pricing

Paid subscriptions move tomorrow from $8 a month and $80 a year to $10 a month and $100 a year. The original deadline was today June 23rd, but the lower rate is staying available through the end of the day for anyone still considering. Subscribing by midnight Pacific tonight means paying the lowest price for the life of the subscription. Anyone who's already a paid subscriber keeps the rate they signed up at, always.

Subscribe now

Beyond the individual subscription

Most paid subscribers join as individuals, but a growing number of subscribers use Card Catalog inside larger contexts: organizations, departments, and teams where the work of evaluating information, using AI well, and doing careful research is part of the job. For anyone who’d like to make the case for employer reimbursement, there’s a request template available. Group subscriptions are also available at 25% off the individual rate for any number of seats a team needs.

Both options, plus full descriptions of every Reference Collection category and what subscribers are saying about the work, are walked through here:

🔑 The Classroom

The Classroom is an expansion of what paid subscribers receive, built so the resources, frameworks, and ideas Card Catalog has been sharing aren’t just things to read once and lose track of. The articles will still come every week, free for everyone who reads them, and the Notes continue too. What’s new is a space where the threads connect: where every resource on a single subject sits together in one view, where an article on source evaluation has a companion lesson that puts the skill into practice, and where paid subscriber content is finally organized so we can find what we want without scrolling through a year of posts to get to it.

This is finally here, and it’s here because of the community that’s been showing up to read, share, question, and build alongside me. Thank you for being part of this, for engaging with the work, and for the trust that made all of it possible.

See you in the stacks!

Subscribe now

This week:

• An AI-powered medical app trained on peer-reviewed journals has been adopted by more than half of American physicians, raising new questions about what happens when doctors and algorithms share the exam room.

• New research from institutions including Stanford and Harvard identified the capabilities AI is least equipped to replace.

• At the G7 Summit, world leaders confronted the reality that access to American AI can be revoked overnight - for reasons no one outside Washington may ever learn.

AI has entered the diagnostic conversation between doctor and patient.

What Happened

A new category of AI tools is arriving in medicine, designed to assist doctors with diagnosis and treatment decisions while they are with a patient. The most prominent is OpenEvidence, an AI-powered app trained exclusively on peer-reviewed medical journals and high-quality research, including content licensed from The New England Journal of Medicine and The Journal of the American Medical Association. More than half of the nation’s physicians now use it regularly, logging 30 million questions and consultations last month alone (nearly double the volume from six months earlier).

OpenEvidence is the front-runner, but the field is filling fast. UpToDate, a longstanding electronic reference for doctors, has added a chatbot interface. Doximity, an online professional network for physicians, acquired an AI company that mines medical literature. Abridge, an AI scribe company, is adding decision-support tools. And OpenAI introduced ChatGPT for Clinicians last month. The competition reflects a shift in what these tools are designed to do: the first wave of medical AI focused on reducing the paperwork that contributes to physician burnout, while this second wave aims to provide clinical guidance at the point of care.

Doctors describe using OpenEvidence as a diagnostic dialogue. They enter the characteristics and symptoms of a patient and receive a summary of the most likely diagnoses along with other diagnoses that are important to rule out, each linked to the underlying research. Dr. Ben Long, a self-described AI skeptic in Corinth, Mississippi, said he now regards the app as “a consultant, a thought partner.” Dr. Barbara Creighton, who treats complex cases at a community hospital in Fairbanks, Alaska, described it as having a team of specialists available on demand, saving her hospital both time and money on outside consulting sessions.

Why It Matters

Medical expertise is being redistributed in real time. A doctor at a community hospital in rural Alaska or Mississippi can now access the same peer-reviewed evidence base as a specialist at a major academic medical center, instantly, through an app on a phone. For a field in which geography has long determined the quality of care a patient receives, that redistribution is a structural change. The gap between what a rural physician can know and what an urban specialist can know has narrowed to the speed of a search query.

The way these tools arrived raises a separate set of questions. Medical technology has historically entered practice through institutional channels: FDA clearance, hospital credentialing, clinical trials, institutional review. OpenEvidence arrived through none of those pathways. Physicians adopted it the way consumers adopt any app, and hospitals learned about it after the fact. Mount Sinai discovered that a third of its 9,000 doctors were already regular users before the institution had any formal relationship with the company. The governance structures that normally vet the tools doctors use were not part of this process, and they are now working to catch up.

What It Means for Us

The information environment in which our doctors operate has changed. When we sit across from a physician, the person on the other side of the conversation may be consulting an AI app in real time, one that draws on the same peer-reviewed research a specialist would reference. The evidence so far suggests this can improve the quality of care, particularly in settings where specialist access is limited. The app is free and available to any government-verified physician in the country, which means there is no barrier between a doctor hearing about OpenEvidence and using it during a patient visit.

The technology is still uneven. Dr. Eric Topol, a cardiologist and executive vice president at Scripps Research, co-authored a paper called “The Illusion of Readiness in Health A.I.,” which found “significant competency gaps” in current AI systems applied to health care. AI has aced standard licensing exams and outperformed doctors in diagnosing certain cases, but it has also failed to accurately summarize research papers and has given wrong answers to diagnostic questions. The tool works best as an input into a doctor’s judgment, and the doctors using it well describe it in exactly those terms.

Bottom Line: OpenEvidence succeeded by training specialized AI on curated, peer-reviewed data rather than the open internet, and by distributing it directly to physicians as a free consumer app. That combination (re: specialized models plus consumer-style adoption) is a template that will likely spread to other professions where expertise, evidence, and high-stakes decisions intersect.

The skills most resistant to AI displacement turn out to be the ones we already use every day.

Read more