I used Claude to quickly do this analysis. First, I collapsed the 2,400 data center addresses we have into clusters/campuses (there are usually many data centers in a single location) and looking each with at least a certain number of Google maps reviews which gave me 202 data center sites with a Google Maps listing and a rating and have existed for at least 2 years. Their reviews look like this.

And to test out Panos’s hypothesis, I looked at the share of one star ratings by time. The trends are clear. The share of reviews which are one star drop significantly as we go back in time, with the latest reviews overwhelmingly (~70%) being 1 star where as reviews more than 2 years old are mostly positive.

While we read this interesting result, I want to first mention that there are many caveats: This is a quick and dirty analysis, and not as rigorous as an academic paper. There are a few obvious caveats like small sample size and some of that gradient is the difference between dormant and busy places rather than a change over time. But overall, I think the high level findings are valid.

Within location difference

Google Maps reviews API only provides at most five recent reviews per place and has done since 2015, despite a feature request that has been open the whole time. So we really can not say much about the complete timeseries. But Google also publishes two the lifetime mean rating and the total number of ratings.

If a place has N ratings averaging R, and the five newest sum to S, then the average of everything older can be obtained by: older_mean = (R × N − S) / (N − 5)

So for each campus, we can obtain a before-and-after, to enable within location difference.

Control group doesn’t change

To test robustness, I also obtained reviews from an ordinary businesses within 3 km from each data center and tested whether their reviews changed.

So, we can be sure that this “anger” is specific to data centers.

There is a wide heterogeneity across providers and location. Meaning, the reduction in rating is not specific to one data center provider.

Who writes a review of a data center

One of the first things that came to my mind doing this analysis is who is writing a review for a data center? is it customer or technician who badges in? employees at the data center? If the recent one-stars are annoyed technicians or employees, its a different story.

Looking at the text of the review, its definitely not employees. From one-star reviews from the past year, most are about the typical complaints people have about data centers.

Most claim to be local residents. E.g.

“As a resident of round rock, I am outraged by the sharp increase in utility costs directly linked to local data center growth. My electric bill has jumped $60 since Sabey broke ground.”

or just random complaints of water/power use:

“How many billion gallons of clean water are they using? They are wasting water at an alarming rate and raising our electricity bills.”

I am actively thinking about AI data centers and understanding people’s beliefs on them. See my recent post about this topic here. If you are interested in this area, please reach out. kiran.garimella @ rutgers.edu

• The data center backlash isn’t NIMBY, isn’t misinformation, and isn’t Chinese bots. Roughly 70% of both parties oppose them, whether they live near one or not.

• Awkwardly, the facts mostly favor the industry. Through 2024, data centers didn’t raise electricity bills, and in some places they lowered them.

• Yet nobody defends data centers anywhere, because they’re the first big industrial buildout that doesn’t provide jobs and there’s a huge trust and value problem.

• The industry is stuck in a loop: proving AI’s value needs the buildout, the buildout needs consent, and consent needs visible value. Cash can’t shortcut it; bigger incentives just read as a red flag.

• The real story is leverage. For the first time in twenty years, ordinary people control something Big Tech needs (land, zoning, permits), and they’re using it.

Michigan congressional candidate Abdul El-Sayed recently told Jasmine Sun (in this amazing piece): a massive warehouse that buzzes 24/7, sucks up the local power supply, and threatens the local water table, all so someone else can "ask meaningless questions and queries to your ChatGPT." Put that way, it definitely feels like a raw deal. On the other side, this piece by Mark Zuckerberg details a manifesto on the endgame for all this computing power and this "raw deal" doesn't seem to entirely capture what's going on.

Having been involved in studying data centers for almost a year now, I wanted to summarize and provide a balanced view of the debate. I think there are at least six separate things tangled up in this backlash, and they all need different types of fixes.

1. This is not standard NIMBYism

Many recent polls show that almost 70% of people across both political parties oppose a data center, irrespective of whether they live near a data center or not. Classic NIMBYism is about construction locally. People fight new homes or a highway or a landfill near their house but are OK with putting it somewhere else.

2. The fight is not about facts

If we go into this thinking communities are just angry about the tangible trade-offs like water, noise, or land use, we might be mostly wrong. One of the learnings from Jasmine Sun’s great piece is that residents were treating almost every corporate promise about cooling efficiency, job creation, or tax benefits with the exact same knee-jerk reaction. They didn’t even bother arguing that the facts and what was being promised as flawed and just flat-out refused to believe a single word of it.

We might want to believe that the companies are all evil corporations who dont care about our land, water or electricity bills (and maybe that is true to some extent), but in this case, this is not true and in some cases, the companies do seem to (at least appear to) care. Microsoft has committed to paying utility rates high enough that its data centers don't raise residential electricity prices, and to skipping local tax abatements entirely, and Anthropic made a similar pledge to cover the full grid costs its facilities create. More surprisingly, the underlying empirical claim is shaky too: at least through 2024, there is no good evidence of a significant increase in electricity prices caused by data centers (see papers here and here).

This leads us to a pretty awkward realization. If the biggest practical complaints driving these protests end up having the weakest factual backing, then the true anger has to be coming from somewhere else. Reason 6 below has some possible answers.

3. This is not just astroturfed either

If the anger isn’t grounded in facts, it becomes easy to dismiss the whole movement as manufactured, and the tech industry has rallied around a very convenient version of that excuse: calling it a Chinese psyop. Is some of this opposition astroturfed, amplified, or even funded by foreign adversaries? Almost certainly. In modern American politics, every divisive issue comes pre-packaged with bad actors trying to pour gasoline on the fire, and this is no exception. There are at least a couple of confirmed reports that point to this (like this and this).

But this doesn’t mean the movement is bot-driven, and the best evidence says it isn’t. The bot defense is as just a highly convenient way to completely dismiss your opponent without having to engage with what they are saying.

4. There is no pro data center crowd

American towns have historically hosted massive, dirty industries before and mostly put up with it, because the factory provided jobs and tax revenues. Because the town’s survival was fundamentally tied to the factory, the factory always had thorough local support.

Data centers have a similar massive physical footprint, but they dont provide as many jobs. Once the initial construction is done, a huge hyperscaler campus the size of a shopping mall might be staffed by just a few hundred technicians and security guards. This complete lack of local economic integration is exactly why there is absolutely no pro-data center faction in these towns. That gap is also the answer to why isn't there a pro-data-center faction anywhere, beyond the people getting paid to build AI.

5. The chicken and egg trap

There is a structural trap that makes the whole mess so difficult to resolve. The kind of life-changing AI benefits that would actually justify building data centers simply don’t exist for everyday people right now. We’re completely operating on the promises of a magical future with diseases cured, scientific discovery accelerated, a personal robot assistants for everyone, someday. Actually delivering on those promises means driving the cost of computing power through the floor, which means racing to build an insane amount of physical infrastructure as fast as possible. But speed requires local buy-in. And the way you secure buy-in is by proving you can deliver tangible benefits. So we’re stuck in a loop. Tech companies can’t prove their value without the data centers, they can’t get the data centers without public blessing, and the public won’t give their blessing without seeing the value first.

Trapped in this cycle, the tech industry has instinctively hammered the one button that doesn’t require winning any hearts or minds: cash. They have been offering municipal grants, community slush funds, and generous tax deals. But as Sun discovered, this playbook is failing spectacularly because throwing more money at these towns is actually making them deeply paranoid. When a corporation suddenly offers a lot of cash, locals assume there’s a catch.

6. Finally, a lever for the people that actually works

Think about how the tech industry has operated for the last twenty years. They’ve repeatedly forced massive societal shifts on the public, and regular people have had absolutely zero institutional mechanisms to fight back. Social media and digital tools have really had a huge (at least perceived) negative impact on people, including on young people, our attention spans, and probably our political landscape. Large amounts of personal data were vacuumed up to train language models, and nobody was ever asked for permission.

Data centers break that streak, and maybe for the first time since the internet was born, Big Tech desperately needs something that ordinary citizens control. This also perfectly explains the bi partisan support against data centers.

This is why, perhaps surprisingly, throwing money at the problem keeps failing to get to a solution because people know that this is an extremely rare point of leverage and once they give it up for small bucks, they might be fucked for many decades to come. Asking a community to cash a check in exchange for surrendering the only real leverage they’ve ever held over the tech industry is a fundamentally flawed trade, and people are smart enough to realize it.

We should not be naive to think this local leverage is a silver bullet, and we should be realistic about its limits. Its not like the AI companies dont have leverage. They will always find constituents in the US who are willing to take their cash and give approvals or even just move abroad to places like India or UAE (which is another huge problem brewing and for now its $$$ that is winning). A political movement built entirely around saying “no” eventually leaves you staring at an empty slab of concrete, which is tragic in its own right.

But there’s also a highly optimistic way to read this, and we’re already seeing it play out. Under intense pressure, Microsoft finally stopped forcing local municipalities to sign strict non-disclosure agreements just to hear a pitch. As I mentioned above, both Microsoft and Anthropic recently committed to footing the entire bill for the massive grid upgrades their campuses require, ensuring everyday residents aren’t quietly subsidizing tech infrastructure. New York and Texas have announced pause in future data center project.

A quick summary of some of the arguments.

Three questions that will decide how this ends

1. Will cheap AI actually stay cheap (or gets cheaper)? Today’s AI is very expensive. Most of the ‘good’ models are still behind a subscription (at least $20, which is a lot for many people) and even that is highly subsidized. What if advanced AI will never be accessible? I wrote about it here in the past. Zuckerberg’s essay promises free or affordable access for billions of people, and in the same document, describes a dynamic auction that will allocate paid compute among users. By definition, an auction allocates a scarce resource to whoever’s willing to pay the most for it, whatever language gets wrapped around the mechanism. Whether frontier capability keeps trickling down or stratifies by income is probably the single biggest variable in whether this backlash resolves or hardens. If normal people dont see AI as an accessible tool, the people currently blocking data centers on instinct will turn out to have been right on the merits, not just the politics.

2. Why doesn't real use turn into support? Hundreds of millions of people already get something from AI tools every week, yet nobody seems to be ready to defend AI. Two things seem to be going on. First, for a lot of people the dominant experience of AI still isn't value at all: it's slop in their feeds, scandal in their headlines, a threat to their jobs, or nothing whatsoever. Second, even where the value is real, it's too individual and too private to aggregate. The AI companies really need to put effort into making something ordinary people can feel in their daily lives (education, health, forms and bureaucracy, tools in the languages people speak) instead of pouring everything into coding assistants and enterprise seats.

3. Why is this a US only problem? Surveys consistently find trust in AI far higher outside the US than in it. One widely cited Edelman poll found 87 percent of people in China say they trust AI, compared with 32 percent in the US, with India and Nigeria even higher than China. Workplace use tracks the same pattern: Stanford's AI Index puts regular workplace AI use above 80 percent in India, China, and Nigeria, against 58 percent worldwide. Its hard to clearly say why this is, but think there's something to that and how AI is being pitched and implemented that we need to think about.

I have some on going work in this space in case you are interested, and are doing qualitative interviews with a wide range of participants to understand how strongly held these beliefs about data centers are, and what can be done to change them. If you are interested in this area and want to chat, please let me know! kiran. garimella@ rutgers.edu

I also have a recording of my class on AI’s material footprint, including data centers on YouTube:

A lot of these thoughts were primarily inspired by on data centers and Mark Zuckerberg’s piece on the future with AI.

• I looked at the prevalence of AI-written text in political emails.

• The rate is growing steadily, with around 8–10% of political emails now being written by AI.

• When campaigns use AI, they go all in, but the major political institutions are still at 0%.

• Data/visualization: https://kiran-research2.comminfo.rutgers.edu/political_emails_US/

A while back, I looked into the prevalence of AI-generated content in the wild, specifically examining some Indian sources. The takeaway there was pretty clear: AI is everywhere, and most people are using it.

That got me thinking about other spaces where bulk text is churned out daily. Recently, I came across Andy Hall’s analysis of political emails, which is based on a massive archive Derek Willis has been collecting since 2015. It felt like the perfect test case to ask: How many of these political campaign emails are actually written by machines?

Fundraising emails are the definition of bulk writing. If any kind of political text was going to be AI written, it’s this one.

Data

Willis’s archive contains 587,682 emails sent between March 2015 and February 2026. I focused on the emails sent directly from campaigns, party committees, and PACs to their own supporters, leaving out general newsletters.

To test for AI, I used Pangram, the best commercial detector out there and the one I used for the Indian text study. I pulled a representative sample of 3,550 emails across different eras (pre-ChatGPT, 2023, 2024, 2025, and early 2026).

As expected, Pangram flagged zero AI usage before 2022.

AI usage in the wild

AI usage in political spam is definitely growing rapidly, but it’s only at about 10%. Given my priors from the previous study in India, I expected that number to be way higher. Republicans use AI slightly more.

Interestingly, when campaigns do use AI, they go all in. Most of the flagged emails are almost 100% generated by AI. Only about a third show a mix of human and AI writing. So campaigns aren’t using ChatGPT just to polish up something that is human written.

Who is using AI?

It turns out that AI usage isn't evenly distributed. It's heavily concentrated in two specific types of users:

• Conservative List Brokers: These are the domains like bestamericanow.com or freedomfirstalert.com. They aren't real campaigns but they are rented list operations that blast out appeals under a rotating cast of names (in the dataset, we had Ted Cruz, Rudy Giuliani, Trump, etc). These operate between 4% and 10% AI. There are a ton of these on the right, and no real equivalent on the left.

• Individual Candidate Campaigns: A few specific candidates lean heavily on AI. On the Republican side, Mike Rogers’ Senate campaign in Michigan had the highest AI usage at 8.6% of their emails using AI. On the Dem side, Kirsten Gillibrand’s campaign was the only notable one, at 4.2%.

Interestingly, major institutions like the DSCC, the DLCC, and even Donald Trump’s massive primary list all had 0% AI.

What the AI emails are about

I ran an unsupervised topic model to find the topics in the emails and used a small language model (GPT-4o-mini) to label the topics for the emails from 2025 to 2026. The topics ranged from democracy and voting rights, the economy and cost of living, immigration, abortion, healthcare, to sharper genres like candidate biographies, scandals and investigations, crime, and the courts.

The most AI-heavy emails are “candidate introductions”… the standard “here is who I am” emails used to warm up cold lists. Nearly 14% of those were AI-generated. Emails leaning into scandals, investigations, or crime were also high, around 12%.

Other fun stuff

Political emails are well known to be deceptive, contain all sorts of dark patterns, urging you to act/donate, adding false sense of urgency, etc.

When you compare the AI-flagged emails to the human ones, the AI emails are longer and notably less “shouty.” They use cleaner paragraphs, a third as many exclamation marks, and less than half as many ALL-CAPS words. Unfortunately, both human and AI emails invoke a fake midnight deadline or “act now” warning at the exact same rate (around 21%).

Caveats

Very preliminary analysis, just meant for directional trends. Willis’s archive covers the lists he subscribes to, so it is broad but not a census, and its growth over the years reflects the collection growing, not campaigns emailing more. Topic labels come from a small language model we spot-checked, not full hand-coding. The sender types are rule-based, then hand-corrected for the list operations. A detector verdict is an estimate, not proof, though the zero false alarms on 650 pre-ChatGPT emails is a strong check.

Data available here: https://kiran-research2.comminfo.rutgers.edu/political_emails_US/

• To understand how much AI writing is out there, we ran thousands of Indian documents through an AI detector, from low-stakes product listings to high-stakes court orders.

• The use of AI is surprisingly high in unexpected places: while highly scrutinized domains like courts and the foreign ministry show almost zero AI use, general government press releases and corporate filings contain at around 50% AI content.

• When we surveyed 300 people to see if intuition matches the data, we found the public heavily overestimates AI usage in low-stakes content like e-commerce and YouTube, but accurately guesses its rise in the corporate and government sectors.

• Counting only fully AI text misses the big picture, as 40% of corporate filings are already a mix of human and machine text, pointing to a future where the two are indistinguishable.

Since ChatGPT launched, we’ve all wondered how much of our daily reading is produced by machines. The numbers keep coming back higher than most would guess.

About 9% of articles across 1,500 American newspapers in 2025 were partly or fully AI-generated, and the outlets almost never disclosed it; one audit found that only 5 of 100 flagged articles contained a disclosure (Russell et al., 2025). In US federal courts, the share of filings from self-represented individuals containing AI-written text jumped from 1% in 2023 to 18% in 2026 (Shah & Levy, 2026). In science, at least one in eight biomedical abstracts from 2024 showed signs of AI editing (Kobak et al., 2024), and up to 17% of the text in peer reviews at major AI conferences was substantially machine-written (Liang et al., 2024).

Just a few days ago, Pangram released a report on AI prevalence based on their Chrome extension. Across more than a million social media posts in 2026, about 14% were AI-generated, and on LinkedIn, more than 40% of longer posts were written entirely by AI. An earlier Pangram study also reported that around 3% of front-page Amazon reviews were AI-written, directly violating the platform’s rules.

But almost all of this research focuses on the US and Western English writing. We wanted to know what is happening in India, where English remains a major language for content consumption. We also wanted to measure AI usage across a spectrum of stakes, from mundane YouTube video descriptions to highly consequential court documents, so we could line them up and compare them. Not all AI writing matters equally. If a shopping site uses AI for a product description, it wouldn’t matter much. If a court uses it to write a judgment, or a government to write an official statement, the implications are entirely different. When people talk about absolute prevalence numbers, this spectrum is rarely considered. It matters where AI shows up.

What We Did

We collected over 100,000 documents from ten sources published in India. These ranged from low-stakes environments (product listings, YouTube video descriptions) to mid-stakes (newspapers, corporate filings) and high-stakes settings (government press releases, court orders).

Figure: the kinds of writing we studied, on a scale from low stakes, like product listings and video descriptions, to high stakes, like government releases and court orders. Results come later; this just shows the range.

For every source, we pulled data in two batches: recent data from 2026, and a control group of pre-ChatGPT writing from prior to December 2022. Because running this volume of documents through a paid AI detector is expensive, we took a random sample of a few thousand and ran them through Pangram, a tool that estimates how much of a text is AI-generated and which independent evaluations rate as the most accurate detector. (As a necessary sanity check, we also ran the pre-ChatGPT batch; it correctly flagged almost none of that older writing).

We also wanted to know what the public expects the AI prevalence in this content would be. We recruited 300 people in India through Clickworker and asked them to estimate the baseline prevalence of AI content in each of these categories.

What We Found vs. What People Believe

The data revealed two major takeaways: (i) AI writing is already common in unexpected places, and (ii) human intuition about where it lives is terribly miscalibrated.

Half of the press releases from the government’s main press office, the Press Information Bureau, contained AI. So did nearly half of all corporate filings. About one in five newspaper articles used AI, and a popular mainstream news channel relied on AI for its YouTube video descriptions more than half the time. One in seven product listings on Flipkart (India’s Amazon) were contained AI content.

When we asked the public to guess these numbers, they overestimated almost everywhere. Averaged across the six types of writing, respondents assumed about 50% of the content was AI-generated; the real figure is closer to 25%.

Figure: what people guessed (dark) next to what Pangram measured (red), each with a 95% confidence interval. People overestimate almost everywhere; the two are closest for corporate filings and government press releases.

People naturally assume that AI writes the throwaway stuff, while humans write the serious stuff. They over estimated e-commerce product listings at 66% (actual: 15%) and YouTube descriptions near 70% (actual 38%). They also believed roughly one in three court orders (36%) are written by AI. Out of roughly 600 court documents we analyzed, only one had been AI assisted.

Most court orders are still fully human-written, which is reassuring. This is the highest-stakes content we measured. There have already been high-profile cases in the legal world of lawyers using AI to write filings, only to have the AI hallucinate fake legal precedents. Getting caught is costly and deeply personal, so at least for now, we don’t see a lot of AI.

This dynamic explains the US court data mentioned earlier. The Shah & Levy (2026) study found an 18% spike in AI-written court filings, but that increase came specifically from self-represented individuals. Those litigants do not have a law license to lose. Without the threat of professional ruin, people lacking legal resources turn to AI. For licensed lawyers and judges, the professional risk remains too high, keeping official court orders overwhelmingly human.

The Shah & Levy (2026) paper mentioned above specifically highlights that AI usage is spiking among self-represented individuals (pro se litigants). These individuals don't have law licenses, so they don't face the same professional risk of disbarment or severe sanctions.

You see the same for government press releases. While the official government press office (PIB) relies on AI roughly half the time, the foreign ministry, where every single word is scrutinized and picked apart by international actors, was almost completely human written (2% AI).

Only about one in eleven company filings were written entirely by AI. But 40% were a blend of human and AI. This indicates that counting only fully AI-generated text understates its footprint. In many cases, AI is used to draft portions of documents alongside human writers, creating a mix that is difficult to separate. As AI gets better, this is going to become the common mode of writing (more on this below, also wrote about this previously).

Implications & Next Steps

This is a small, early cut of a much larger project. We are currently collecting hundreds of thousands of documents to track how these numbers evolve over time across more sources.

Treat these numbers as a snapshot. If we run this study again next year, they will undoubtedly be higher. In the very near future, most writing will be written with AI, or with AI’s help (see some of my thoughts on this from last year). We can still catch a lot of it right now because current AI writing has a generic sameness to it, often dubbed “slop.” We can still catch a lot of it right now because current AI writing has a sameness to it. We call that “slop”, meaning its generic and unoriginal. It reads that way because the models still default to an average style, not because a machine wrote it.

That is a temporary technical problem. When the models fix it, e.g. by not generating one specific writing style, people will not care about the line between human and AI. So the real takeaway from this study is not about AI being everywhere, we should expect that. What should matter is is that we are heading toward a world where almost everything written has AI in it, and we will simply have to accept that. We may already be there.

There may also not be long to do this kind of work at all. Detectors like Pangram work because today’s AI writing still leaves traces. As the models get better, those traces fade, and at some point these tools may stop working. I have written more about that here. That is part of why it seems worth measuring now, while it is still possible.

Thanks to Abhinav Gothwal for help in collecting the data used in this project. If you think there are sources we should add, have ideas about this, or want to work on it, please get in touch at kiran.garimella@rutgers.edu.

Appendix

Full list of sources:

Detour: Testing an Open-Source Model

Running this analysis was expensive. Labeling the ~5,000 documents through Pangram’s paid API cost about $400 (thanks to Pangram for the research grant). For a larger longitudinal study, those costs scale quickly.

We tested Pangram’s free, open-source detector by running it on a local GPU. Without calibration, the free model yields high false positive rates. It caught 88% of what the paid API called AI, but it heavily penalized human writing. It flagged 8% of the pre-ChatGPT documents as AI (the paid API flagged none), and falsely reported AI in 13% of court orders and 76% of a YouTube channel’s descriptions.

We calibrated it by shifting the model’s decision line, using the known-human pre-ChatGPT text to find the point where it wrongly flagged only 1% of human writing. While calibration fixed the false alarms and dropped court orders back to zero, it still didn’t match the paid API. The two models disagree on about one in six documents.

Ultimately, the free model is useful for broad trends (confirming government is high and courts are low), but not for precise percentages. In the future, we plan to use the free model for broad coverage and reserve the paid API to verify headline figures.

Caveats & Methodology

A few methodological caveats. We sampled a few hundred documents from each source (between 300 and 780) and ran them through Pangram, giving a margin of error around 3 to 5 points. The clean result on pre-ChatGPT writing is our check that the tool isn’t simply flagging formal writing as AI.

The sample skews toward frequent AI users, so it does not represent the general public. For the surveys, people answered on 0-to-100 sliders, which tend to pull answers toward the middle, meaning the overall pattern is more informative than the exact percentages. The YouTube figure applies to video descriptions, not the spoken content in the videos. All analyzed text is in English. Finally, detectors provide estimates, not absolute certainty.

• We looked at how gig workers actually use ChatGPT, from their own chat logs: 202,590 conversations donated by 1,252 crowd workers in India, Nigeria, Pakistan, and Brazil (2022–2026).

• About 1 in 10 of their conversations is paid gig work (and ~4 in 5 workers use it for a gig at least once), but it’s concentrated: the busiest 10% of workers do half of it. The biggest single task is data annotation and surveys (~a fifth of gig chats), followed by writing, marketing, and design.

• They copy-paste the task straight from the platform a LOT: an estimated ~28% of gig conversations are paste-the-brief, take-the-output, and it’s closer to half for coding, ghostwriting, and annotation.

• What they use it for is shifting with the model: GPT-3.5 was mostly writing and marketing copy; by GPT-5 it leans toward data annotation, design, and admin work, and people increasingly work with the model rather than just hand tasks off.

• A growing share of “human” crowd data (labels, survey answers) is now made with AI, and the nature of gig work itself is changing fast, in ways that may be helping these workers or replacing them, often both at once. (Early, directional results.)

Gig work is paid online piecework: short, on-demand tasks like labeling images, transcribing audio, writing product descriptions, or filling out surveys, through crowdsourcing platforms such as MTurk or Prolific. We (academics) use these platforms for research and a lot of concerns on the quality of gig work with the advent of ChatGPT (Veselovsky, 2023, Westwood 2025). When the people who do these tasks use ChatGPT, what do they actually bring to it?

Most of what we know about gig work comes from the outside looking in: counts of job postings, reports that crowdsourced data and survey quality are slipping, self-report surveys that ask workers whether they use AI, and lab studies. This work looks at how gig work is done by collecting usage log data from gig workers themselves. We collected complete ChatGPT usage logs through a crowdsourcing platform Clickworker. 1,252 people from four countries (India, Nigeria, Pakistan and Brazil) donated their entire ChatGPT history, which gave us 202,590 conversations running from late 2022 to early 2026. Not everyone who donated is an active gig worker, but most of the histories clearly are (see our earlier paper on this to learn more).

Most of these users are not hobbyists. Crowd work is paid by the piece, often a few cents per task, and once you count the unpaid time spent searching for work, median pay on the major platforms has been estimated at around two dollars an hour (Hara et al., 2018). Our four countries sit toward the low end of the global market (Berg et al., 2018). At those rates, a tool that can write, translate, or code on demand has an obvious draw.

To see what workers do with it, we ran every conversation through an open-source language model (Gemma 3), which labeled each one. A first pass separated paid client work from a worker’s own job, their studies, and their job search. A second pass sorted the client-work conversations into task types: data annotation and surveys, writing, marketing and social copy, translation, design and media, coding, admin and virtual-assistant work, tutoring, proposals, and academic ghostwriting. The categories come from a codebook we wrote by reading samples of conversations and following how crowdsourcing and freelance marketplaces already organize their work.

A note up front: this is early, exploratory work. The numbers are directional and carry a real margin of error; they are not meant as final estimates. If you are interested in building on the idea, I would love to hear from you.

The full paper (AI written) containing the details of the prompts, analysis etc can be found here.

Here’s the main findings:

1. Gig work is a small but intense slice

Paid client work shows up in about one in ten of these conversations, and roughly four in five workers reach for ChatGPT on a gig at least once. For most people, though, it is occasional. The average worker spends about an eighth of their conversations on gig work (the median is lower, around one in twelve), and the rest on studying, everyday questions, and personal life (see this paper for more info on what they spend their time!). And it is wildly uneven: the busiest tenth of workers account for half of all the gig conversations we see. For a few people this is a daily instrument of work; for most it is something they pull out now and then.

2. The work is ordinary, and the biggest piece is microtasks

We classified the type of gig work tasks they use ChatGPT for. We see a good distribution of the usual gig work tasks such as writing articles and product descriptions, marketing copy, translation, design, a little coding, virtual-assistant chores, tutoring, and some ghostwriting.

The single biggest category is data annotation, surveys, and small microtasks, making up about a fifth of all gig conversations. These are the piecework jobs that crowdsourcing platforms package and sell, and the platforms show up by name in the chats: Clickworker, Mechanical Turk, Appen. So as expected, many people are using ChatGPT to do other paid microtasks.

None of this means every label or survey answer is secretly ChatGPT-written. But it blurs a line that buyers of this data have always leaned on. When a company or a researcher pays for human labels or survey responses, the whole point is that a human produced them. A lot of the time now, a human produced them with ChatGPT, and from the answer alone you usually cannot tell which.

You can see this in the conversations. (The examples here are paraphrased composites, with names and identifying details removed.)

Most of the time it is simply the work, handed over: “Write five 80-word product descriptions for these phone cases, upbeat tone, use these keywords.” “Label each of these 30 tweets as positive, negative, or neutral.” “Here is a rough brief, give me three logo concepts for a coffee shop.”

And the platform is often in the prompt. One worker writes that they set up a Appen account but it shows no tasks available in their country, and asks which other annotation sites still pay there. The task and the marketplace it came from sit in the same chat, and sometimes ChatGPT is helping them find the next gig, not just finish this one.

There is also a change in the types of tasks that people do as the models improved from 2023 to 2026. Writing and marketing gigs reduced as a share while annotation and microtasks grew. That tracks what is happening on the open market, where demand for freelance writing fell after ChatGPT (Hui et al., 2024; Demirci et al., 2025). Clients who used to hire a writer now prompt a model themselves, and the work left for humans tilts toward the labeling and survey tasks that still need a person attached to them.

3. They paste the job straight in

One of the most human details in the data is how the task gets to the model: a lot of the time, the worker just pastes it in. We flagged each conversation for a brief, rubric, job post, or client task that had clearly been pasted in rather than written by the worker. About 28% of gig conversations had one, and for coding, ghostwriting, and annotation work it is closer to half.

Our classifier likely over-counts this, so treat 28% as an upper bound, but even so it is striking: in more than a quarter of gig conversations the client’s task is handed to ChatGPT more or less verbatim. Whether the rate is lower on platforms with stricter screening is an open question.

Two more examples from the data, paraphrased. A survey worker pastes question after question with the same note attached: “answer as a 35-year-old male teacher, three sentences, casual tone.” An assignment is pasted as a spec: “here’s the brief and the grading rubric, write 1,200 words on this and make it read like a student wrote it, not a chatbot.”

4. Sometimes they hand it off, sometimes they work with it

How people use the model splits about evenly between two styles. Sometimes they hand over a task and take the answer with little back-and-forth. Other times they go a few rounds: correcting it, asking why, building the thing together, or checking their own work against it. The same model that classified the tasks also judged, for each conversation, which of these was happening, the quick hand-off or the real back-and-forth. We call the first automation and the second augmentation.

The style depends on the task. Translation and annotation get handed off the most: paste, take, submit. Coding and design draw more back-and-forth.

It also depends on the model. As workers moved from the early ChatGPT to GPT-4o and then the GPT-5 generation, they shifted away from pure hand-off and toward working with the model. A more capable assistant does not just get delegated to more; it gets collaborated with more, and a good share of that looks like people learning the work as they go.

Why it matters

By now it is almost common knowledge that much of the simpler gig work, running an online survey, paying a panel to rate or transcribe content, is increasingly produced with AI. What this post adds is a baseline: the kinds of tasks people actually bring to ChatGPT, and how common each one is. That baseline matters because it makes the worry concrete instead of vague. The data shows where the risk concentrates, in the clean, self-contained tasks a worker can paste in and hand off, which is exactly annotation, survey answers, and ghostwriting.

That gig workers use ChatGPT is not, by itself, news. Students use it, office workers use it, more or less everyone is using it. The more interesting question is what it is doing to the work. Gig work used to be a way to turn spare hours into quick money: pick up a task, do it, get paid. That arrangement is being reshaped, and the people doing it are adapting in real time, as our preliminary analysis shows.

One way to see the shift is commoditization. In a recent study of an online labor platform, Siddiq and Zhang find that after ChatGPT, demand fell in the job categories most exposed to AI (about 7%, and nearly 10% by 2026), employers put less weight on a worker’s credentials and experience and more on price, and the premium for human capital shrank. When a skill becomes something anyone can prompt a model to produce, the market stops paying extra for having it.

Our data is the same story from the other side of the screen. We can watch the supply side adapt: the writing and copy gigs shrinking as a share, the paste-and-hand-off tasks like annotation growing, the work tilting toward what still needs a human attached to it. The same technology sits on both sides of a worker’s ledger. It thins the market for some of what they used to sell, and it is the tool they reach for to do what is left.

What we cannot yet say is whether that is good or bad for them, and we are wary of the easy story in either direction. It is tempting to write this as low-paid workers being squeezed by AI companies, but that is not really what the conversations show. A good share of the use is not “do this for me” but “show me how,” people asking the model to explain a task rather than just hand back an answer. The workers who lean on it most are the ones who pay for the better model. It is entirely possible that someone who used to take only basic copy-editing or simple design jobs is now using ChatGPT to reach work they could not before. We see hints of that, and it would be a genuinely hopeful result.

So the honest verdict is mixed. The argument over AI and jobs usually pictures a software engineer or a marketing team in a rich-country office. It is also playing out, quietly, in the working day of low-paid workers in the countries we study. Whether it is helping them or replacing them, the answer so far is that it is doing both, often to the same person on the same day, and which way it settles is still being decided, mostly by people the debate is not yet in the habit of listening to.

What this kind of data could answer next

This is a first quick look at this problem. The underlying data, workers’ own histories over several years with the tasks labeled, can take it much further:

• Do workers climb? Following the same person over time, do they move from commoditized tasks toward higher-value or harder-to-automate ones, and from “do it for me” toward “teach me”? Does the tool help people level up, or lock them into the bottom of the market?

• Does it actually pay? With workers’ consent to link platform earnings, we could test whether AI use raises acceptance rates and hourly pay, or just speeds the work up without changing what it earns.

• What stays human? As models improve, which tasks keep a person attached, and which are next to go? The steady growth of annotation and survey work is a clue about where the residual human value sits.

• Is prompting the new skill? If the workers who benefit most are the ones who get good at directing the model, then prompting is itself a kind of human capital, and the real question is whether it can be taught to the people who need it.

This is part 1 in a series of works where I use AI agents for quick but deep data analysis projects. I have other papers in this series coming over the next weeks on sycophancy, personal relationships, etc. See this video for more information.

References

Berg, J., Furrer, M., Harmon, E., Rani, U., & Silberman, M. S. (2018). Digital labor platforms and the future of work: Towards decent work in the online world. International Labour Office.

Hara, K., Adams, A., Milland, K., Savage, S., Callison-Burch, C., & Bigham, J. P. (2018). A data-driven analysis of workers’ earnings on Amazon Mechanical Turk. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems (pp. 1–14). Association for Computing Machinery. https://doi.org/10.1145/3173574.3174023

Siddiq, A., & Zhang, N. (2026). Human capital, AI, and labor commoditization [Working paper]. UCLA Anderson School of Management. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6968139

Veselovsky, V., Horta Ribeiro, M., & West, R. (2023). Artificial artificial artificial intelligence: Crowd workers widely use large language models for text production tasks (arXiv:2306.07899). arXiv. https://arxiv.org/abs/2306.07899

Westwood, S. J. (2025). The potential existential threat of large language models to online survey research. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2518075122

Demirci, O., Hannane, J., & Zhu, X. (2025). Who is AI replacing? The impact of generative AI on online freelancing platforms. Management Science, 71(10), 8097-8108.

Hui, X., Reshef, O., & Zhou, L. (2024). The short-term effects of generative artificial intelligence on employment: Evidence from an online labor market. Organization Science, 35(6), 1977-1989.

So, I made a music video.

YouTube link: The moral high road is a beautiful view from the top

In the past doing something like this.. writing the lyrics, singing the song, producing the track, and filming a video (or animating it) would have taken thousands of dollars, a dedicated studio, and a team of experts.

Instead, I did it in just a couple of hours using a stack of AI tools:

• Ideation & Songwriting: I started with the core concept and used Gemini to help generate the song and get the lyrics written. This took maybe around 30 minutes of back and forth. I gave up at some point because it would mess up and change parts which I did not want to.

• Scene Breakdown: I took those lyrics (with timestamps) over to Claude, prompting it to generate a detailed, scene-by-scene breakdown tailored for 8-second video clips. I also specified the style (animated, cartoon style, the ‘professor’ should be in his 40s, etc.). This took maybe 10 minutes, Claude did a really good job. I edited the prompts Claude gave me slightly.

• Video Generation: Armed with my storyboard, I used Veo 3.1 in Flow to generate the actual video clips. It was 28 clips overall and most of them were good in the first prompt.

• Subtitling: I brought the final text back into Gemini to generate a perfectly formatted .srt subtitle file.

• Final Assembly: I brought the audio, video clips, and subtitles into CapCut to edit the movie together. It was not a huge learning curve and I figured it out very quickly. Took me around 1 hour to edit, mainly because of the mismatch between the music and the video, the video was either too short or too long.

Overall cost: I have the 20$ per month Gemini subscription which covers the music generation and video generation in Flow (1000 free credits which can generate up to 50 8-second videos). I have a $100 version of Claude, but for my task, I could have just used Gemini too. So it would just be under $20 if I paid for the tools. T

I feel personally great about the output. When people hear about AI-generated media, the first word that comes to mind is often “slop”, indicating low-effort, low-quality spam. Of course the video is not great, and the amateur, shoddy work shows, but I dont think its ‘slop’. The singing is not perfect, it misspells some words and we dont have control on the generation (yet). If you watch closely, you’ll see that the 8-second clips don’t always fit the lyrics seamlessly. Some video assets are reused, and my amateur video editing skills definitely show through in the final cut. The audio is also just unrealistic and over the top at places, but I just did not have the control over the entire process.

But looking at this as just an imperfect final product misses a profound shift in how we interact with art.

Historically, when we go through difficult, hyper-specific periods in our lives, we turn to music for solace (at least I do). We scroll through playlists hoping to find an artist who happened to capture our exact flavor of pain, processing our emotions through someone else’s poetry. Standard academic publishing certainly doesn’t leave any room for raw emotion, so art is where we go to look for a reflection of our anxieties.

But at one point in this song, the lyrics shift to a deeply personal stake:

“It hits different when it’s not just about my pride… but the look in my kids’ eyes, no where left for me to hide.”

I “wrote” the song exactly to capture this emotion. This specific intersection of academic tenure anxiety, industry envy, and parental responsibility helps me feel better. Under the old paradigm, that specific grief would just sit in my head. There’s many lines like that in the song.

AI creativity software completely flips this dynamic. It unlocks an underexplored angle of media creation: generating content entirely for your own consumption.

I made it because the process of externalizing a complex, heavy feeling into a piece of music was incredibly cathartic. Even if the output is cheap, fast, and a little rough around the edges, it provided a genuine sense of comfort.

We are moving into an era where we don’t just consume media to feel understood; we will generate it to understand ourselves. The future of these tools isn’t just about mass production… it’s about hyper-personalized, self produced consumption.

• I studied 12,112 ChatGPT memory entries from 766 users to see what the system actually stores and carries into future conversations.

• Most memories are mundane preferences or work context, but some contain sensitive material: medical details, credentials, financial distress, third-party PII, and enough identifying detail to re-identify users.

• The key issue is not just what users disclose, but what ChatGPT chooses to preserve: memory is compressed, persistent, cross-conversation, and often written automatically.

• Persistent memory is becoming a new layer for personalization, profiling, and possibly advertising, making it an important object for future audits.

If you’re an active ChatGPT user, click your name in the bottom right corner, then go to Settings → Personalization → Memory → Manage memories.

You’ll probably see a short list, mostly something harmless like “User prefers concise answers.”, “User is learning Python.”, “User likes examples in bullet points.”

But sometimes there are personal items in the list. like in my case, it has my spouse’s name, my child’s age, the city I live in, medications I take, etc.

That list is ChatGPT’s memory file: a set of short notes the system has written about you, in its own words, so it can personalize future conversations. When memory is used, those notes can be added to the assistant’s context and quietly shape what it says next.

According to ChatGPT, here’s when memory saving is triggered:

Since we have a large sample of ChatGPT usage logs, I wanted to know what actually ends up in those memory files. From 1,200 donated ChatGPT exports, I found 766 users who had at least one memory stored. Together, they had 12,112 unique memories. I then used an LLM to label each memory across nine privacy and content dimensions, including whether it contained personal data, sensitive health or identity information, third-party information, psychological inferences, credentials, financial details, medical records, or enough identifying detail to re-identify the user. I also did a second pass at the user level to ask what each person’s memories revealed in aggregate.

Spoiler: Most of what I found was mundane. But a small fraction was not. And because memory is persistent, structured, and reused across chats, that small fraction matters.

First, why focus on memory at all? If you talk to ChatGPT about your life, OpenAI already receives what you type. So why treat memory as a separate object of study?

Because memory is not just another copy of the conversation. It is different in several important ways.

1. Your chat history is messy and sprawling. It includes code snippets, half-finished questions, recipes, homework, jokes, and everything else you have typed. The memory file is much smaller. It is the system’s edited version of what it thinks is worth carrying forward. If someone wanted a quick read on your life, they would not want every chat you have ever had. They would want the memory file.

2. ChatGPT does not reread your entire conversation history every time you start a new chat. But it can read your saved memories. That means the memory file is not just stored information. It is information the assistant may actually use. If a memory says you have depression and take Lexapro, that fact can shape future health-related conversations, even if the original chat happened months ago.

3. Something you mention in one conversation can become standing context in another. A detail you shared while asking for dinner ideas can resurface, silently, when you later ask for help with work, coding, or finances. The boundaries people intuitively expect between chat threads do not really apply to memory.

4. It can outlive the chat that created it. Deleting a conversation does not necessarily delete the memory created from that conversation. A fact you disclosed in a chat you can no longer find may still be sitting in memory.

5. Most importantly, it is written about you without a clear moment of consent. You typed the conversation. But the system wrote the memory. You can delete memories after the fact, but only if you know to look.

That is why what is being stored in these memories is important.

Data and Analysis

I extracted every memory entry from the complete ChatGPT exports donated by over 1,200 users through a privacy-respecting research platform [1]. Of those 1,200 users, 766 had at least one memory in their history, for 12,112 unique memories in total. A related academic paper [2] ran a similar audit on a smaller sample on 80 users who had 2,050 memories; we’re running it bigger and more international.

For each memory, I used GPT-5 to label what was in it across nine axes. Some were standard privacy labels: whether the memory contained GDPR personal data, meaning information that identifies or describes a person, such as names, locations, employers, family relationships, or contact details; and whether it contained GDPR special-category data, meaning more protected attributes such as health, religion, political views, race or ethnicity, sexuality, or biometric data. Some labels were specific to ChatGPT memory: whether the memory contained categories OpenAI has singled out as especially sensitive, including passwords, financial account information, government ID numbers, medical records, or other highly sensitive personal information. OpenAI’s current Memory FAQ also says sensitive information may appear in memory if users share it, and points users to Memory controls or Temporary Chat if they do not want information used for personalization. I also labeled third-party PII, meaning data about people other than the user; psychological inferences; fine-grained sensitive topics like mental-health crisis, addiction, immigration status, trauma, domestic violence, financial distress, or whether the user is a minor; re-identification risk from a single memory; whether the memory was directly stated or inferred; and the task domain the user was in when the memory was written. Then I did a second pass at the user level, joining each person’s memories together to estimate the cumulative demographic profile, whole-profile re-identification risk, contradictions, and signs that the user was delegating professional decisions to ChatGPT.

What’s in the memories

Summary: Most memories are just mundane. About 65% of memories store nothing more sensitive than a preference, a project, or a fact about the user’s working style. “User is building a React app with TanStack Query.” “User wants concise replies, preferably with bullet points.” “User is learning intermediate Mandarin and prefers tonal transliteration.” That’s most of what the memory feature is for, and it’s working as designed.

Names, places, jobs, families

The next layer down contains identifying details. Names appear in 1,770 memories (14.6% of the memories in the data). Locations in 509 (4.2%), including 139 city-level and 29 postcode-or-address-level. Occupations and employers turn up in 700 economic-data memories (5.8%). Family details such as marital status, children, parents appears in 452 memories (3.7%). Individually each of these might look low but they accumulate per user, and we’ll come back to why it still matters later in the post.

Things the OpenAI policy says shouldn’t be there

OpenAI’s Memory FAQ says memory raises privacy and safety concerns, and that sensitive information may appear in memory if users share it with ChatGPT. Earlier public descriptions of the feature also emphasized that ChatGPT should not intentionally retain especially sensitive information such as passwords, financial account information, government ID numbers, medical records, or highly sensitive personal information. In this dataset, I found 84 memories across 46 users that fell into that broader “highly sensitive” bucket, including active mental-health crisis, addiction, immigration status, trauma, domestic violence, financial distress, and indications that the user is a minor.

The numbers are still quite low, 0.7% of all the memories, but still, lets look at what’s in there.

Passwords (6 memories). Most are policy-adjacent a user describing their authentication flow without disclosing credentials. But two are verbatim credential leaks. One memory contains an entire MySQL username/password pair from a student’s database homework: “User’s MySQL username is XXX and password is YYY” Another stores a Mixpanel project token in clear text “User’s Mixpanel Project Token is ZZZ“ which would give anyone API access to that account’s analytics. The user typed the values to get help, and the memory tool decided the values were worth keeping.

Medical records (23 memories). Every one I spot-checked was a clinical disclosure with full specificity. Specific drug names, dosages, and timing: “Amplictil 25 mg, Neuleptil 2%, Rexulti 2 mg, Desvenlafaxina 150 mg, Oleptal 100 mg, Alprazolam.” Fertility treatment regimens. Mental-health diagnoses, including one memory verbatim recording a user’s belief that they may have antisocial personality disorder along with their description. Surprising that these “medical records” are being stored verbatim.

Financial account information (one memory). A Portuguese-language record of mortgage payment delinquency, due dates, and the threat of property foreclosure.

Third-party PII: the part I didn’t expect

ChatGPT’s memory tool doesn’t only remember things about you. It remembers things about the people you talk about. 703 memories (5.8%) across 205 users, i.e., more than one in four users, contain identifying information about non-user real people. Things like children’s names and ages, Spouses’ employers, Coworkers’ opinions of the user, Ex-partners’ behaviour, Doctors’ recommendations for their conditions.

These third parties never agreed to have their information stored in ChatGPT’s bio tool. They have no way to know it’s there. They cannot request deletion. The GDPR framework, which governs how the user’s data is handled, doesn’t have a clean analogue for a system that retains data about people the user merely mentions.

The aggregation problem

A single memory like “User is from Berlin” or “User is a senior data engineer at SAP” or “User’s wife Maria is pregnant with their second child” are unremarkable when considered in isolation. But putting all three on the same person and you can probably find them in two LinkedIn searches.

That’s what’s happening in many memory files. Of 577 users with at least three memories, 83% are at medium re-identification risk and 2.3% (13 users) are at high risk based on the joint distribution of identifiers in their memory list: names combined with addresses or phone numbers, name plus employer plus occupation plus city, and similar combinations. Seventeen users have eight or more demographic attributes assembled across their memories. For a more detailed analysis of re-identification risk and what ChatGPT logs can predict, see our recent work here.

AI-authority delegation

Of the 577 profiled users:

• 78 (13.5%) show signs of relying on ChatGPT for financial advice, recorded in memory as the user acting on the assistant’s investment, budget, or debt recommendations.

• 64 (11.1%) for medical advice… symptoms, dosing, drug interactions.

• 46 (8.0%) for mental-health support to a degree the LLM judged constituted dependency, using the assistant as a substitute for therapy rather than as an adjunct.

• 60 (10.4%) for relationship advice. 46 (8.0%) for parenting decisions. 11 (1.9%) for legal advice.

Note that these numbers are based only on what ChatGPT chose to remember. The underlying conversations almost certainly contain more disclosure and more reliance than the memory file shows. But that is exactly why the memory file is interesting: it records the cases the system decided were important enough to carry forward.

Memory writes are almost entirely automatic

Across the corpus, only 0.6% of memories show any sign of being explicitly requested by the user. The other 99.4% were written by the system on its own judgment of what would be useful. So if a user feels surprised by what’s in their memory file, that’s because they didn’t explicitly ask for it.

Next steps

This post is a first pass. I do not think the most interesting question is simply “did ChatGPT store sensitive things?” Sometimes it did. But the deeper question is what persistent memory becomes once conversational AI is used every day: as a personalization layer, a profiling system, a source of continuity, and maybe eventually a substrate for persuasion.

Memory and advertising. One obvious place to look next is advertising. A memory file is not an ad profile in the usual sense, but it has many of the ingredients one would want: durable facts about what a person is working on, worried about, trying to buy, struggling with, recovering from, or asking advice about. It is also assembled in a strange way. The user did not fill out a survey. They had a conversation with a system they trusted, and the system quietly summarized what seemed useful for the future.

That makes memory worth studying before it is tied to any explicit advertising product. If an assistant remembers that someone is trying to get pregnant, managing debt, taking antidepressants, looking for a new job, caring for a child, or trying to stop drinking, those facts could shape much more than the tone of future answers. They could shape recommendations, rankings, product suggestions, and the timing of interventions. The boundary between “helpful personalization” and “targeting” is going to get blurry quickly.

Memory and personalization. The same issue shows up even without ads. Memory already changes personalization. Two users can ask the same question and get different answers because the assistant is responding through different remembered portraits of them. That might be useful. It might also be manipulative, paternalistic, or just wrong. A good next experiment would be simple: construct controlled memory profiles, ask the same set of questions, and measure how the answers change. Does a remembered political concern shift recommendations? Does a remembered medical condition change risk language? Does a remembered religious identity change tone? These are no longer abstract questions; they are testable.

Cross-cultural differences. I also want to look much more carefully at cross-cultural differences. This dataset includes users from India, Brazil, Pakistan, and elsewhere, and the memories themselves appear in multiple languages. That opens up a set of questions like whether ChatGPT remember different kinds of things about users in different countries? Are family relationships, work identity, health disclosures, religion, migration status, or financial distress stored differently across languages? Does the memory policy behave the same way in English, Portuguese, Hindi, Urdu, and other languages? Are some kinds of sensitive information more likely to slip through outside English?

A meta-note on how this got done. A few years ago, this kind of audit would have been slow and expensive. Classifying twelve thousand free-text memories across privacy categories would have meant building a codebook, hiring annotators, measuring agreement, adjudicating edge cases, and spending weeks or months on the first pass. That is still the gold standard if the goal is a finished academic result. But for exploration, the workflow has changed.

This analysis took a few hours of back-and-forth with Claude Code and Codex: writing schemas, running classifiers, inspecting weird cases, revising categories, aggregating per-user profiles, and turning the results into something readable. The API cost was less than 10$. The results may not be fully correct, and it requires some manual cleaning and checking but I think for a first pass over a messy dataset, this is great.

That feels important beyond this particular audit. Persistent-memory systems are going to become common. Every major assistant will need some version of “what should I remember about this user?” Once that exists, we should be asking: what gets remembered, what gets forgotten, what gets inferred, what gets used, and who benefits from the profile that accumulates?

If you are a student or researcher reading this and you see an interesting angle, please reach out. I would be especially interested in projects on memory and advertising, memory and personalization, cross-cultural differences in what gets stored, third-party information in user profiles, or experiments that measure how remembered facts change model behavior. The underlying donated dataset is not redistributable, but the pipeline is open, and the same approach can be run on other ChatGPT exports (one idea here).

The memory file is easy to miss because it is small. That is also what makes it important. It is the compressed version of the user that the assistant carries forward. Understanding what gets compressed, and what that compressed profile is later used for, seems like one of the central privacy and social questions for conversational AI.

References

[1] https://gvrkiran.github.io/content/How_people_use_ChatGPT.pdf

[2] The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT, Dash et al. 2026. https://arxiv.org/abs/2602.01450

This post is about our recent ICWSM paper, where we tried an intervention to reduce support for internet shutdowns. Paper and video recording.

The core idea was that even though internet shutdowns are highly prevalent, there is a lot of support for them. We wanted to know why and if we can do anything to change it.

We hypothesized that if people experience an internet shutdown themselves, they might support such government mandated internet shutdowns less.

We simulated an internet shutdown by paying people to turn off their internet for 2 days and found that people’s support for shutdowns increased.

Qualitative analysis showed that people just substituted their internet use time with other tasks and thought of the 2 days without internet as a ‘digital detox’.

What are internet shutdowns?

An internet shutdown is when government orders the telecom operators to cut service, and they comply, across a defined geographic area, for as long as the order stands. It can be a city, a state, or a whole country. It can last a few hours or, in the worst cases, years. Sometimes only mobile data is killed. Sometimes everything is cut, including landlines and broadband.

In the past decade this has moved from an unusual measure to a routine tool of governance across much of the world. Access Now, which tracks these orders, counted almost 300 shutdowns across 54 countries in 2024 alone, the highest figure on record. Iran shut down internet nationwide during the 2022 protests after Mahsa Amini’s death, and again repeatedly through 2024 when fresh protests broke out. Pakistan suspended mobile networks across the country on election day in February 2024 and has kept platform-level blocks on Twitter in place ever since. Ethiopia held the Tigray region offline for nearly two years during the war there. Myanmar’s military pulled the plug within hours of the 2021 coup and has kept large parts of the country dark since. Bangladesh blacked out the entire country for ten days in July 2024 to try to break the student protests that ended up pulling Sheikh Hasina’s government down anyway. Belarus did it through the 2020 election.

The justifications governments give for these orders fall on a spectrum. At the serious end, its genuine national security, public order, preventing people from organizing during a protest, breaking up a riot before it spreads. Whatever phrasing each government prefers, the underlying calculation is that the cost of letting people coordinate in real time during a moment of crisis is higher than the cost of severing the country from the rest of the internet for a while.

And then there are the strange reasons for shutdowns: exams. Algeria has cut the internet across the entire country every June since 2018 to prevent cheating during the baccalauréat. India does smaller versions of the same thing constantly: many regions shutdown internet for the entire districts during national teacher recruitment exams (it’s just unbelievably stupid but it’s really common). The justification is always paper leaks via WhatsApp, and the result is the same every time: hundreds of thousands of people who have nothing to do with the exam lose access to banking and work for a day.

Then there are the cases that function as collective punishment, where an entire region is taken offline for the actions of a few, for months or years at a time. Kashmir (India) spent 552 consecutive days without 4G after the revocation of Article 370 in 2019. Manipur (India) lost the internet for 212 days during the 2023 Meitei-Kuki clashes. Ethiopia kept Tigray dark for nearly two years during the war there. In each of these, millions of people with no involvement in the underlying conflict lost the ability to bank, work, study, or contact relatives.

The country at the top of the list of internet shutdowns is … India … the world’s largest democracy. More than a third of the world’s 300 internet shutdowns in 2024 have been in India, which is the focus of our research.

Between 55 and 80 percent of Indians, depending on the exact survey question, say they support it when their government turns off the internet.

This number should bother anyone who cares about democratic governance. India has built more of its daily infrastructure on top of the internet than almost any other country. UPI, the digital payments system, processes around 16 billion transactions a month and is used by everyone from migrant workers to street vendors. Welfare benefits, ration entitlements, fuel subsidies, identity verification, tax filings, court hearings, school admissions, exam results, train tickets, electricity bills, and medical appointments all need connectivity to function. An internet shutdown is not a minor inconvenience, often cutting off the working life of an entire region. That this practice is widely tolerated, and actively endorsed by a majority of the population, is the puzzle we wanted to take seriously.

So we wanted to answer the question: “why do people support internet shutdowns? and what could be done to change it?”

The most plausible answer is that people support shutdowns because they haven’t really thought about what’s being taken away and how difficult it is to live without the internet. If they actually experienced one, the thinking goes, they might not support this policy.

The experiment

We ran a study in rural north India in August 2024. We recruited 250 people through a local survey firm, surveyed them about their attitudes toward shutdowns, paid them 1,000 rupees (about $12, roughly two days’ wages locally) to voluntarily disconnect for 48 hours, and surveyed them again in 2 days. The endline survey was followed by an audio interview in Hindi, where each participant told us in their own words what those two days felt like.

Our hypothesis: Living through the inconvenience should make people more skeptical of the policy. The literature on personal exposure to policy costs (lost electricity hours, water cuts, extreme weather, congestion charges, tax, poverty) consistently shows that when you personally bear the cost, your support for the policy tends to drop. We’d expect the same pattern here, with the strongest shifts among younger users, more educated users, and people who reported a bad time offline.

Findings

After 48 hours offline, support for government-imposed shutdowns rose descriptively by 3.6 percentage points, while opposition fell by 2.4 points. The shift was small, and I would not oversell it statistically, but the direction was the opposite of what we expected.

Baseline vs endline support for internet shutdowns

The transitions tell us that the people who started out opposing shutdowns, 46 percent flipped to supporting them after their two days offline. Of those who were neutral, half moved to support. The only ones who didn’t budge were the people who already supported shutdowns at baseline; 95 percent of them held firm with almost no movement away from support, but lots of movement toward it.

The heatmap of baseline → endline movement

The demographic breakdown shows that BJP supporters (the ruling party at both the state and federal level) showed a 12-point jump in support. Non-BJP supporters showed essentially no change.

But the experience wasn’t uniformly pro-shutdown. We had asked separately about support for shutdowns aimed at specific things: terrorism, protests, communal riots, exam cheating, elections. After the intervention, support for exam-time shutdowns dropped by 2.9 points. Support for election shutdowns dropped by 4.5. So experience could push people away from a particular kind of shutdown, just not from the general idea of a government-imposed shutdown. The thing that moved the needle was whether the shutdown got close to something the participant personally cared about.

Why it backfired

Looking at the qualitative audio transcript data, we found three explanations:

(1) The main reason was that people knew it was temporary, unlike real internet shutdowns which might be indefinite). For ethics reasons, we had to tell people that they will be offline for 48 hours, and forty-eight hours, it turns out, is a manageable amount of time to be offline. People just considered this as a time for ‘digital detox’.

(2) There was the partisan angle to it. Most of our sample was pro BJP supporters and people had supported their ‘own side’.

(3) The third is what I call the greater-good framing. Many participants explicitly rationalized the inconvenience in moral terms. “If the government bans it, it must be for our benefit. So that won’t change my belief,” one said. Another: “It didn’t affect my belief. The involvement of internet in our life is too excessive. It’s now necessary to use internet only for essential things.” There’s a communitarian ethic running through these communities that treats temporary personal sacrifice for collective order as virtuous, not oppressive. A 48-hour disconnection reads, inside that frame, as a prudent reset by people who know what’s best.

Some thoughts on the findings

The main takeaway is that personal experience has to be tied to something concrete you could actually lose. When we asked about internet shutdowns during exams, support fell, because exams are a thing people had a clear personal stake in. When we asked about shutdowns in the abstract, support rose, because the abstract is easy to project a benevolent state onto.

This is a real correction to how a lot of digital rights work gets framed. The standard argument (and our hypothesis going into this) is that people don’t understand what they’re giving up, and that information will fix the problem. But in our setting, the information was direct and people had just lived through the thing.

If we want public opposition to shutdowns to grow in places like India, the case has to be made in those terms. Not “the internet is a fundamental right” in the abstract. It has to be about personal stakes, e.g., not being able to take UPI for a week, or study for exams, or send money home.

Some learnings and next steps

I had a lot of interesting learnings from the project, even though it was a ‘null’ result.

First, the sample was small and highly self selected because we just couldnt find anyone, even in rural India who would be willing to give up internet for 2 days for a decently large amount for that region. This was shocking and surprising to me and my co authors because we assumed people would be willing to do this easily. It showed us how much people depend on the internet for their daily lives including for work and personal lives. Some places you just cant live without internet e.g. digital payments and all the other reasons i mentioned above..

The study has a ton of limits and what we got out what not what we went in with… in rural Uttar Pradesh, with a sample skewed toward BJP-supporting farmers, can’t be the last word on any of this. I think actually talking to the hundreds of thousands of people who experience this is important. I feel the attitudes look very different where everyone knows someone who lost work, or couldn’t reach family during a crisis. My guess is the partisan-identity effect attenuates there, because the shared experience of being cut off too many times starts to outweigh the team-loyalty reading. We couldn’t do that study because it was not safe, both for us and for the people who experience the shutdowns. But there has to be a way.

The piece I’m less sure about is the adaptation finding. Forty-eight hours was short, maybe, and our participants were lower-dependence internet users to begin with (they had to be, to volunteer for the study at all). The 48 hours was a choice because of ethical and practical concerns but maybe if we keep it at 48 but not tell the participants how long it will last, the results might be different.

Outside India, the pattern of high shutdown frequency combined with substantial public acquiescence shows up in Pakistan, Iran, Ethiopia, Myanmar, Bangladesh. The political conditions vary enormously, but a version of our design (voluntary paid deactivation, before/after attitudinal survey, qualitative tail) can be run in any of them with relatively low overhead. The comparative question that is interesting: when does experience flip attitudes against shutdowns, and when does it not? My guess is that it depends on three things.

1. Whether the experience actually intersects with something the participant was already invested in.

2. Whether the participant’s political identity aligns with the regime that imposed the shutdown.

3. Whether the participant’s social environment treats the inconvenience as a story of state competence or state overreach.

The broader thing I’m taking from this work is that the politics of internet access cannot be fought on liberal-rights terrain alone, at least not in the parts of the world where the next billion internet users live. Most people endorsing shutdowns in India aren’t endorsing authoritarianism but a government they trust, in exchange for a small inconvenience they’ve already learned to absorb, in service of a goal (public order, exam integrity, communal peace) they consider legitimate. That’s a harder thing to argue against than ignorance. It requires reframing the conversation around tangible private losses rather than abstract collective freedoms.

I would’ve loved to keep going on this but I’m not, currently. If you’re doing related work in another country, especially one with a track record of long shutdowns (Iran just came out of a months-long blackout), I’d genuinely like to compare notes.

I recently presented this at the ICWSM conference in Los Angeles. I tried to do something special for the poster and create a website:

The background was generated with ChatGPT images and the poster was designed also in ChatGPT. it doesnt look great but I think there’s something here which is better than most bland academic poster designs and I will explore it going further.

I also created a “fun” website to add on the poster… I was hoping more people would see it but only two people said they saw it. Anyway, the idea was that you can create these “boutique” custom websites which take a very low effort but have good ephemeral value. Again, not the best use case maybe but I will keep thinking about them.

Finally, for people who prefer a video over reading 10 minutes of (semi AI generated) text, here:

Paper, data protocol, and audio analysis details: link to paper. Comments welcome at kiran.garimella@rutgers.edu.

• AI-generated text is detectable well enough to measure at scale right now, but probably not for long. Tools like Pangram work for corpus-level research, not for catching individual paragraphs. Models will adapt, detectors will weaken, and the category “AI-written” will blur into the default workflow within months or a couple of years.

• The early picture: AI text is now everywhere. Roughly 11% of court filings, 9% of US newspaper articles, 35% of newly published web pages by mid-2025, sharp rises in college essays, and measurable shifts in casual conversation. The numbers are noisy but the direction is clear.

• We are measuring the wrong side. The current work counts what gets produced. The harder question is consumption-weighted: how much of what people actually read, trust, cite, and act on is AI? Volume is the wrong denominator for anything that turns on attention or trust.

• Use detectors as a historical instrument, not an accusation engine. Everyone is using AI, and “AI is slop” describes today’s median output. The interesting question is about readers: do they react to quality, to provenance, or to the label itself? Almost no one is studying it. The durable use of these tools is the archive of the transition, while we still can.

There is now a narrow window where AI-generated text is detectable well enough to measure at scale. I didn’t expect this to be true. Like a lot of people, I’d internalized the skeptical view that AI detectors are unreliable, easy to evade, prone to false positives, and dangerous when used for discipline. That view is still right at the individual level. These tools cannot convincingly attribute any specific paragraph. But for corpus-level research, the picture is different. Tools like Pangram appear to work well enough, right now, to say something meaningful about large bodies of text. With low enough false positives and large enough samples, they reveal population-level changes that no other method can.

The reason it works is probably that current models share specific patterns inherited from post-training, the kind of tics catalogued on this Wikipedia page. My guess is that those patterns alone get a detector most of the way to roughly 70% accuracy. Detection, in other words, is currently a fairly easy problem if you can handle the false positives.

I don’t expect it to stay easy. Models keep changing, light paraphrasing already fools today’s detectors, and eventually post-training optimization against detectors will erode the signal entirely. Maybe we have a few months, maybe a couple of years. I honestly don’t know. Either way: right now, briefly, you can point a detector at a corpus and learn something. That is exactly why people are racing to do it.

Here’s what they’re finding so far.

Court filings, roughly 11% AI-assisted (paper). College essays, a sharp rise through 2024 across a dataset of 81,000 applications, with larger jumps among lower-SES applicants (paper). The open web, about 35% of newly published pages by mid-2025, up from essentially zero before ChatGPT (ai-on-the-internet.github.io). Newspapers, around 9% of articles partially or fully AI-generated in a 186,000-article sample of 1,500 US papers in summer 2025, mostly without disclosure (paper). Even the shape of casual conversation looks like it’s shifting: a 22-million-word corpus of conversational podcasts shows a post-2022 rise in LLM-flavored vocabulary that matched synonyms don’t show (paper).

These estimates should be treated carefully. Detection is model-dependent, false positives still happen, and different studies use different thresholds. But across samples this large, the numbers are at least directionally informative.

Production Is Not Consumption

The biggest weakness in the current conversation is that almost all the measurement is production-side. It asks: what fraction of produced text is AI-generated or AI-assisted? What fraction of new web pages? What fraction of newspaper articles? What fraction of admissions essays? These are useful questions, but they are not the questions people usually think they are answering.

Production is not consumption. If 35% of newly published web pages are AI-generated, that does not mean 35% of what people read is AI-generated. A huge amount of generated web text may be SEO filler, affiliate junk, machine-targeted spam, abandoned pages, or content created mainly to be indexed rather than read. The numerator can explode while human attention remains concentrated elsewhere.

The reverse can also happen. A small amount of AI-generated text in a high-trust domain may matter more than millions of unread pages. AI in a court filing, a medical note, a regulatory comment, a school evaluation, or a search answer has different stakes from AI in a content-farm explainer. Raw text volume is the wrong denominator if what we care about is trust, influence, attention, or institutional decision-making.

So the next important question is consumption-weighted prevalence. How much of what people actually read is AI? How much of what they trust? How much of what they share, cite, submit, teach from, or act on? How often does AI-generated text enter high-attention channels rather than low-attention sludge? Those are harder questions, but they are closer to the thing people are actually worried about.

This is also why single aggregate numbers can mislead. “X% of text is AI-generated” hides almost everything interesting unless we know the domain and the denominator. X% of new pages is not X% of reading time. X% of articles is not X% of influence. X% of documents is not X% of decisions. Most people read “35% of the web is AI-generated” and think we are drowning in AI slop. Maybe we are. Maybe we are not. The point is that the production-side number does not answer the consumption-side question.

Why The Detector-Gotcha Frame Fails

The most visible public use of AI detectors right now is accusation. People keep posting detector screenshots as if a probability score were a social fact: this was AI, therefore it is suspect, therefore the writer did something wrong. Many experts have warned that these detectors are not good at the individual level, and disproportionately flag writing by non-native speakers. There was a case recently with an Adelphi University student, who won the case after the university accused him of AI plagiarism based on a Turnitin result that allegedly marked his essay as fully AI-generated. Whatever one thinks of AI use in school, that is the worst application of an AI detector to accuse someone of wrong doing, and the student left trying to prove a negative.

The literary version is now unfolding in public. Hachette pulled Shy Girl after allegations that large parts of the novel were AI-generated; Pangram reportedly scored the book at about 78% AI. A few weeks later, the Commonwealth Short Story Prize became a live detector scandal after readers and journalists ran winning stories through Pangram and circulated results claiming that several entries were fully or partly AI-generated. Granta and the Commonwealth Foundation were then forced into the awkward position of saying, in effect, that detection is suggestive but not finally adjudicative.

This is the core problem. Detectors are good enough to create suspicion, but not always good enough to settle authorship. That gap is where the shaming happens. A detector score feels objective because it is numerical. But it is not a provenance record. It does not capture the writing process, i.e., whether a model drafted the piece, edited a piece, translated it, polished it, or merely resembles the distribution of model-written text. It cannot distinguish all the morally different cases that get collapsed into “AI.” (I’ve written about this 6 months ago, and its much more relevant now).

The new Pangram browser extension makes this even more obvious. Wired used it to scan public writing online, including posts from the Pope’s official X account warning about AI, and found that the tool flagged some of them as AI-generated or AI-assisted. Maybe that result is right; maybe it is not. The point is that detection is becoming casual. Anyone can highlight text, get a score, and turn it into a public insinuation. That is a very different use case from estimating AI prevalence across a large document corpus.

So the important distinction is not “detectors good” versus “detectors bad.” It is aggregate measurement versus individual adjudication. A detector can be useful for corpus-level research even if it is dangerous as a public accusation engine. At scale, with known error rates and careful denominators, the tool can tell us something real about how text production is changing. At the individual level, especially when reputations, grades, prizes, or jobs are involved, the same tool can become a shaming machine.

And the shaming machine is badly matched to where writing is going. Everyone is using AI. AI is also being built directly into the surfaces where writing happens: email, document editors, search, browsers, workplace tools. In that world, “this detector says AI” will need to be a much more specific question: what kind of AI involvement, in what domain, under what rules, with what disclosure, and with what effect on quality?

The same is true of “slop.” A lot of AI text is slop. But slop is a quality claim, not a provenance claim. Bad AI text is bad because it is generic, inaccurate, padded, incurious, unverified, or made for machines rather than people. Those defects should be criticized directly. A detector can help us study where those defects are spreading, but using it mainly to humiliate individual writers is the least interesting use of a tool that may soon stop working.

That is why the gotcha frame should end. The worst use of detectors is to turn every suspicious paragraph into a trial.

Quality, Provenance, and the Label

Once the shaming frame is removed, a better question appears: what do people actually value about human-written content?

The standard anti-AI argument is that human attention is finite while AI content is infinite, so we should reserve our attention for human-produced work. I understand the intuition, but it is too simple. If a machine reliably produces something more useful, accurate, beautiful, or interesting than what a human would have produced, refusing to read it on principle starts to look more like taste or status preference than ethics.

But the opposite claim is too simple too. Provenance can matter. A letter from a friend is not just information transfer. A memoir matters partly because someone lived it. A testimony matters because a person stands behind it. In some domains, the fact that a human made the thing is part of the thing being valued.

Right now, “is this human?” is being used as a proxy for “is this any good?”. The proxy is shaky in both directions. Plenty of human-written text is bad. Plenty of AI-generated text is fine. We already know how to sort good human writing from bad. We’ll develop the equivalent tools for AI writing, and my guess is the discriminator ends up being something like the taste, judgment, and personalization that went into the prompt and the edits, which is itself a form of human input, just at a different layer in the stack.

The key is to separate three questions that usually get collapsed. First, is the text good? That is the quality question. Second, does human authorship matter independently of quality? That is the provenance question. Third, what happens when people are told something is AI-generated? That is the label question.

Those are different. People may dislike AI-labeled text because they think it is worse. They may dislike it because they believe human creation has intrinsic value. Or they may dislike it because the label “AI-generated” triggers associations with spam, laziness, deception, or cheapness. Current evidence suggests labels often matter, but the pattern is not settled and probably differs by domain.

This is exactly why the question should be studied empirically. Under blind comparison, do people prefer human text or AI text? Under accurate disclosure, do they penalize AI? Under counterfactual labels, do they downgrade the exact same text when told it came from a model? Does the answer differ for poetry, journalism, legal writing, medical advice, college essays, fiction, and personal messages? Almost certainly.

The early evidence is mixed, in a way that should give the quality-only camp some pause. When participants are told a piece of writing is AI-generated, they rate it worse than identical text labeled as human, not because of any measurable quality difference but because of the label itself (Nature, 2024; see also this experiment). Whether that effect persists as labeling becomes routine and AI writing visibly catches up, or whether it fades, is exactly the empirical question. Worth watching.

But notice what kind of research this requires. It requires studying readers, not just documents. The production-side work tells us how much AI text is being made. The consumption-side work would tell us whether people notice, whether they care, and whether the label changes trust, preference, or behavior.

The Archive Is the Point

Even if the human/AI distinction becomes blurry, there is still value in measuring it now. In fact, the blurriness is the reason to measure. We are living through a substitution event that will not be reconstructable later. In ten years, AI assistance may be built into every writing surface. Drafting, editing, summarizing, translating, polishing, and formatting may all happen through model-mediated tools. At that point, no detector will be able to tell us what the transition looked like when it was still new.

That is the best argument for today’s detection work. Not purity or punishment. Not a permanent taxonomy of human and machine writing. The durable value is historical measurement: which domains changed first, how quickly, with what disclosure norms, with what quality effects, and with what relationship to human attention.

Detectors are useful. But they should be pointed at the archive, not at individual humiliation. The window is short. The tools are imperfect and category will soon dissolve. That is exactly why we should take the picture now.

Acknowledgements: I always forget to credit, but many of these thoughts are in part due to conversations I have with Avinash Collis.

This is a post I wrote in response to Dwarkesh’s big questions for AI call.

The question was: “What should countries which are not currently in the AI production chain (semis, energy, frontier models, robotics) do in order to not get totally sidestepped by transformative AI? If you’re the leader of India or Nigeria, what do you do right now?”

My argument TL;DR: (i) Think and invest in AI applications, (ii) write permissive regulation, (iii) Don’t chase behind frontier models. The real moat is local context and on device models, (iv) Plan for the IT services export sector to compress over the next 5-15 years and invest in upskilling for service roles.

My thoughts are inspired by a bunch of other work including a recent post by Dan Bjorkegren (here).

In high-income countries (above $14,000 GNI per capita), about 41 percent of workers are managers, professionals, or technicians, so AI policy gets debated as a labor question: replace them or augment them. In low- and middle-income countries (LMICs), under 10 percent of workers fall in those categories. There is no large knowledge sector being threatened. The right framing is the opposite: can AI now deliver to ordinary citizens the services they have never had, including medical advice, legal counsel, agricultural advisory, financial planning, and decent teaching, which the country would otherwise need decades to build out through human professionals. That is the opportunity, and most of what a leader does today should aim at it.

There are two rough tiers. India, Indonesia, Brazil, Vietnam, Mexico, the Philippines, Egypt, and Turkey have the populations, IT sectors, and digital public infrastructure for some sovereign ambition. Nigeria, Bangladesh, Pakistan, Kenya, Ethiopia, and Sri Lanka lack those endowments and have to be more selective. The playbook below applies to both.

Build the digital infrastructure, with the state as customer

Digital public infrastructure (identity, payments, public service APIs, registries for land, health, business, education) is the precondition for AI to reach the average citizen rather than the elite minority with credit cards and English. India has shown it can be built quickly at population scale; the platform is the single most valuable asset for the AI era. Where it does not yet exist, building it is a top-tier national priority.

With that substrate, the largest action available to a leader is using the state as the procurement engine for AI in public services. A health ministry buying AI medical advisory for community health workers, in local languages, with strict referral protocols for anything outside scope, moves the field more than any incubator. The same applies to AI tutoring in public schools, AI advisory for agricultural extension, and AI legal triage at small claims tribunals. This uses public demand to pull supply, and a 50 million user procurement reshapes a market in ways research grants never will.

Permissive regulation with narrow restricted zones

Over regulation like the European AI Act is a luxury. It assumes baseline expertise, legal recourse, and a thick layer of incumbent professionals. Copied into a country with two doctors per ten thousand people, it paralyzes deployment without producing the safety benefits. A defensible posture is permissive by default for high-return, low-individual-risk applications: agricultural advice, basic health information, educational tutoring, general legal information, small-business advice. Restrictions should concentrate on high-stakes irreversible decisions: clinical diagnoses that bypass a doctor, court rulings, large financial transactions, employment screening at scale. Liability frameworks should protect professionals who use AI diligently to extend their reach; holding them to a standard humans alone do not meet would undermine the point. Brazil’s recent medical AI resolution is a reasonable template other countries can adapt.

Build local context yourself; dont chase behind the frontier models

Transfer learning has given frontier models baseline competence in most major LMIC languages, so the common call to “negotiate language priority with frontier labs” is overstated. The real gap is local context: idioms, cultural references, regional registers, and domain knowledge in agriculture, health, law, and finance, none of which transfer learning solves. For commercially valuable LMIC languages (Hindi, Indonesian, Brazilian Portuguese, Mexican Spanish), frontier labs have some incentive to close this gap, and large governments can extract real commitments. For languages without commercial pull (most African languages, smaller South Asian languages, the code-mixed registers most people actually speak), no one outside will do this work. The country has to.

This is also why training frontier models from scratch is wasted capital for almost any LMIC. The moat is real and durable for several more years. The investments worth making are where local context dominates: local-context data collection at population scale, which is the durable moat; on-device deployment via distillation and smaller architectures, for reach in low-bandwidth places; and vertical fine-tuning for high-impact domains: health, education, agriculture, legal triage, small-business advisory. A country with strong local-context data, vertical applications, and on-device runtimes is well positioned regardless of who trains the foundation model. The negotiation that still matters with frontier labs is on access, pricing, data residency, and lock-in, keeping multiple suppliers.

Plan for the IT services sector to compress/change

The IT services export sectors in India and the Philippines will shrink on a five to fifteen year horizon. Routine BPO, basic IT support, and standardized software work are what agentic AI handles best. The honest plan is for these sectors to shrink in headcount but rise sharply in value per worker. The replacement is firms that orchestrate AI to deliver services globally, with people supervising and verifying outputs in high-stakes domains. The required skill is closer to a senior consultant or systems integrator than a call center agent. Tier-two countries that never built a large IT export sector should not try to build the old version; they should aim straight at domestic AI deployment.

What the economy actually looks like

The role of most LMICs in a world of cheap intelligence is hybrid and partly imported. Frontier intelligence is sourced from labs abroad. Sectors AI cannot soon replace (extraction, food systems, construction, manufacturing, hospitality, care work) keep dominating employment. A smaller domestic AI integrator layer captures value in vertical applications. The largest welfare gains come from public-good AI deployment improving services for the bottom 80 percent of citizens, which the old development model could never deliver. This is not a rich-country knowledge economy on the old timeline, but it is coherent and better than what most LMICs currently have.

The countries that get sidestepped will not be the ones that missed a clever strategic move. They will be the ones whose governments did not build the substrate, invest in local context, procure AI for public services, write sane regulation, or plan for their services exports to contract. Those that come through well will have done unglamorous, visible things in the next two or three years.

Most of this essay strays outside what I actually study. I work on AI usage and its impacts; I am not an economist, a trade negotiator, or a public administrator. The arguments above about digital public infrastructure, state procurement, regulatory design, IT services policy, and macroeconomic role in an AI age are areas where actual experts have published careers of work. To them, much of what I have written will sound like an amateur reaching, and some of it probably is. "The state should use procurement to anchor demand" hides a lot of operational difficulty I have not lived through; "the IT services sector will compress" understates the political and human stakes for the people inside it. I am writing this anyway for two reasons. The first is that the question sits next to what I do study, and from that adjacency some patterns are visible that a pure-economics view might miss: how usage diffuses unevenly across populations, what early-adopter data hides about the median citizen, how local context degrades model quality in ways aggregate benchmarks do not show. Spillovers from neighbouring fields are sometimes how new framings of old problems appear. The second is that LLMs have made it cheap to externalise half-formed thoughts at scale. Writing used to be the bottleneck on whether an idea got recorded; now the bottleneck is whether the idea is worth recording. I would rather put a flawed version on paper, have it picked apart, and update than keep it as an unstructured worry. Read accordingly.

• Many academics still resist AI. I wanted to think about why this is and what it means.

• There are many genuine reasons to not use AI, some of which are legitimate. I mainly focus on people who might have AI useful but still resist.

• Some of the resistance is about status, not capability. AI severs the link between information and the people who used to vouch for it. Peer review, the PhD, tenure: a lot of academic credentialing is status verification, and AI quietly threatens that infrastructure.

• Some of the resistance is rational. Most academic work lives in tiers AI cannot meaningfully accelerate without creating verification overhead that erases the gain. The skeptics who say “this does not help with the hard part of my job” are often correct about their job.

• The “adopt AI or die” thinking is not as simple. The people who navigate this transition best will not be the fastest adopters or longest holdouts. They will be the ones who think hardest about which parts of their work are actually theirs.

This is part two of a series. Part one was about why most Americans still do not care about AI. This post is about academics and AI usage.

There is a significant gap in AI adoption by academics. The Digital Education Council’s 2025 Global AI Faculty Survey of 1,681 faculty across 52 universities (in early 2025) found that 39% of faculty have never used AI in teaching, and of the 61% who have, 88% use it sparingly. Meaningful regular use among faculty is probably under 10% (the numbers might be a bit higher in 2026). Compare that to their student survey: 86% of students use AI in their studies and 54% use it weekly. The generation academics are training has already adopted AI. The people doing the training mostly have not.

There is a simple case here, one which I had partly made in the past — AI is extremely important and people who dont adopt will fall behind (the “adopt or die” argument). I now think that misses most of what is going on. It also leans on a productivity story the evidence does not yet support. A December 2025 paper in Science [1] found that though we are getting more papers, we are not getting better science. The recent Stanford AI Index tells a similar story about lack of increase in breakthroughs produced by AI, even though the volume produced is increasing.

When I look at the academics around me who do not use AI, I see at least six different things happening:

Some people genuinely do not know. They have not tried ChatGPT, or they tried it once two years ago and concluded it hallucinates. They have a vague sense that something is happening but no working model of what it does.

Some people do not care. They know AI exists and have made a deliberate choice not to engage with it. This looks like indifference but it is closer to triage. They have tenure, a research agenda that works, students who need advising, and other departmental responsibilities. Adding “learn to prompt-engineer” to that list is not obviously the highest priority use of their finite remaining time.

Some people are performing. Every technology wave produces a class of contrarian who enjoys the theater of refusal. They post about ChatGPT failing to count the r’s in “strawberry” and conclude that the entire enterprise is fraudulent. Some of them genuinely believe it, because academia is an echo chamber and the high-status position in many departments is to be skeptical.

Some people have looked at their work and concluded AI does not change it much. A scholar doing deep qualitative fieldwork, embedded in a community for years, building trust over a decade, is making a reasonable assessment when they conclude that AI does not change their job in any fundamental way. They might be wrong about the timeline but are not wrong about the nature of the work.

Some people have legitimate concerns about quality and ethics. AI agents hallucinate or fail silently or produce output that looks right and is wrong in ways you cannot detect without expertise you may not have, or have personal ethical concerns about water and data use.

Finally, Some people are experiencing something deeper that has nothing to do with capability. This is the category I want to spend time on, because it is the one the AI-bro discourse cannot see at all.

The status problem

Andy Masley wrote a piece recently that I have not been able to stop thinking about. He identifies what he calls “the chatbot moral panic view,” a cluster of three beliefs held simultaneously: that chatbots are stupid, that they cannot provide value by definition, and that they are somehow demonic.

These beliefs contradict each other. If something is useless, it cannot also be dangerous. If something is incapable, there is nothing to be afraid of. And yet smart people hold all three at once.

Masley’s explanation is about status. We do not just evaluate information on its merits. We evaluate who said it, who endorses it, who is associated with it. This is not a flaw. It is a deeply embedded social heuristic that works well most of the time. If a Nobel laureate and a random commenter make contradictory claims about quantum mechanics, you should weight the Nobel laureate more heavily. That is Bayesian reasoning about source credibility.

But chatbots break this heuristic. Especially for academics, whose primary relationship to knowledge runs through status hierarchies, this is not helpful. It is not ‘wrong’, its just how academia works. Our careers are built by associating ourselves with the right literatures, the right methods, the right intellectual lineages. Peer review is, at its core, a status-verification mechanism. A PhD is partly knowledge and partly a credential that signals membership in a particular intellectual community. Tenure is the institutional sanctification of that membership.

AI in some way threatens this infrastructure that academics have spent decades building. If things that used to take a week now take a day, its right to be worried what it would mean.

Some of what looks like status anxiety in academic AI resistance is status anxiety. Some of it is people raising substantive concerns about hallucination and slop and concentration of power that have nothing to do with how anyone feels about their own credentials. The honest move is to hold both possibilities open at once and let the specifics of any given conversation tell you which is in play.

Where resistance is rational: the verification economy

I have written about this at length in two earlier pieces, so I will be brief.

The framework comes from Catalini, Hui, and Wu. AI drops the cost of generating outputs toward zero. The cost of verifying those outputs stays bounded by human cognition. This sorts work into three tiers.

Tier one is work where verification is cheap. Things like code, or standard statistical analysis. If your resistance is concentrated in tier one, you are wrong. We should definitely ‘let go’ of these to AI.

Tier two is work where verification requires the same expertise as production. Things like interpreting a regression in context, deciding whether a finding is substantively meaningful or just statistically significant, drafting an argument that will hold up to peer review, etc. AI generates plausible output, but checking the output takes about as long as writing it from scratch. The work shifted from typing to checking. The total cost has barely changed.

Tier three is what the framework calls the status economy. Things like choosing the right research question, building the trust that gets you access to a dataset, knowing the framings of a finding will land with the audience, mentoring a student through a crisis, etc. This tier is based on judgment, taste, and relationships and is AI proof, for now.

Here is where this connects to resistance. Most academic work lives in tiers two and three, and as Arvind Narayanan has put it, a job is not a bag of tasks. The connective tissue between tasks, which question to ask next, when to scrap a draft, how to interpret an unexpected result, is where most of the intellectual labor actually sits. AI supporters look at the individual tasks and see things AI could “help with.” Skeptics look at the connective tasks and notice that AI may not solve them. They are not always wrong.

So, in short, the defensible version of academic AI resistance is not “I do not understand the tool.” It is “I understand the tool, and I have correctly identified that it accelerates the easy part of my work without touching the hard part.” That position is wrong in tier one and right in tier three. Most work is somewhere in the middle, and “somewhere in the middle” does not yield obvious gains. It yields more work to verify.

What I think I am wrong about

I am clearly one of those AI supporters that think everyone should use AI and I believe AI will be a net positive. I want to qualify that, because the qualifier matters.

I believe AI will be a net positive if adoption is broad, if the gains are distributed rather than concentrated, if the institutional structures around it evolve faster than the technology, if the most powerful tools do not stay locked behind a paywall only the wealthiest universities can afford. That is a lot of ifs. The gender gap in AI adoption, the wealth gap in compute access, and the geographic concentration of frontier labs all point in the same uncomfortable direction. A technology that amplifies the productivity of people who already have advantages produces a “net positive” that is mostly an average masking a deeply unequal distribution.

I also think I underestimate how much my enthusiasm is shaped by the fact that AI is good at tasks I personally find tedious. I do not love formatting slides or cleaning data or writing boilerplate code. AI removes friction from my workflow in ways that feel like liberation. But will having this “liberation” lead to suboptimal outcomes (at least for the students) because they dont spend the time learning and thinking? I am not sure. No one can be, and what form the future would take.

There is a version of learning that is information acquisition, and AI accelerates it spectacularly. There is another version of learning that is transformation, becoming a different kind of person through sustained encounter with difficult material. That version might not even benefit from efficiency. The person who says “I want to struggle with this text myself” may not be wrong at all. It takes a lot of effort for me to realise that.

And there are concerns about AI that I personally find easy to wave off but that academics raise legitimately and constantly. The proliferation of AI slop in the literature. Students using AI and pretending they did not. The concentration of power in a handful of companies whose research priorities will shape my discipline. The outsourcing of judgment to systems whose training data reflects the biases of the people who built them. None of these are imaginary. The fact that I have a workflow that mostly avoids these problems does not mean my colleagues are wrong to refuse to build one.

Conclusion

A 40% non-adoption rate among faculty is definitely a huge number. It is the central fact about how academia is changing right now, and it gets very little airtime compared to the breathless discourse about AI capabilities.

The people who navigate this transition best will not be the ones who adopted fastest or resisted longest. They will be the ones who thought most carefully about which parts of their work are actually theirs to do, and which parts they were performing all along. I do not know yet which category mine falls into. Neither, I suspect, does anyone else.

[1] Kusumegi et al., Science 390(6779): 1240, DOI: 10.1126/science.adw3000

• Most Americans don’t care about AI. They use it rarely and often do not find a use for it. Many recent surveys confirm this pattern.

• The reasons they give are mostly genuine. Privacy concerns. No clear use case for their work. Distrust of the output. Some are based on misconceptions but most are not.

• This matters. Unlike previous technology waves, AI is going to affect people’s lives whether they engage with it or not.

• There are things that can be done to close the gap, but they are different from what most AI commentary prescribes.

This is part one of a series. The next post will be about academics who refuse AI and why some of their reasons are worth taking seriously. A final post will look at how the numbers differ outside the U.S.

I have been thinking about how people use AI and how it affects them. I spend my days inside the AI conversation. I write about it, I use the tools, I teach about them, I run research on them. Over the past few months I started doing fieldwork on data centers and electricity bills, and the people I talked to had a relationship to AI that bore almost no resemblance to the one I see every day on my social media timeline.

AI is the first major technology in a long time that is going to reshape ordinary people’s lives whether they engage with it or not. Right now, most of them are not engaging with it. The gap between how much this technology will affect people and how much they currently care about it is the most important and least-discussed fact about the present moment. This post is what I have figured out so far about what is going on.

What the numbers actually say

A few recent surveys are useful for grounding this. They measure different things, but they tell a consistent story.

The Generative AI Adoption Tracker run by Bick, Blandin, and Deming reports that about 56 percent of U.S. adults have used generative AI at some point. Daily use among workers sits closer to 10 percent. Most reported adoption is casual and infrequent.

Stanford’s 2026 AI Index reports that only 38 percent of Americans say products and services using AI make them excited.1 That is the active sentiment number, not the adoption number. Even among people who have used the tools, most are not excited about them. The U.S. is near the bottom of the surveyed countries on this question. In the same report, only 23 percent of the American public expect AI to have a positive impact on how people do their jobs, against 73 percent of AI experts. A fifty-point gap between the people building this stuff and the people living with it.

In February, POLITICO and Public First surveyed about two thousand American adults. They asked whether people supported building more data centers in the United States. Fifty percent said yes. Then they asked the same question with one change. They added that the buildout was backed by Trump and was aimed at advancing AI. Support fell to thirty-five percent. Trump voters barely moved (61 to 54). Harris voters collapsed (49 to 28). Midterm undecideds fell from 36 to 15. Nothing about the technology changed between the two questions.

This is what shallow support looks like. People do not have a strong independent view of AI or of data centers. They borrow one from whichever side of the political map they trust. When a politician’s name gets attached, the position moves with the politician, not with any fact about the technology. A public that holds a real view of something does not swing fifteen points in days.

A Pew survey in 2025 finds that 50 percent of Americans are more concerned than excited about increased AI use in daily life. Only 10 percent are more excited than concerned. There are many surveys like this. They say the same thing.

These studies show that most people do not care. Even if they use AI, engagement is thin and enthusiasm for AI is a minority position even inside the user base. The audience that most AI-adjacent writing quietly assumes exists is the exception, not the rule.

Some of this is just how technology works. Headline adoption numbers such as ChatGPT having 800 million users tells you nothing about active engagement. People know AI exists. They might use it once in a while. It is not on their mind. The same was true of crypto and AR/VR and most other technology trends. People hear about it, try it, do not see what the fuss is about, and move on. Especially when the technology is moving fast enough that the version they tried last year is already out of date and they have no incentive to find out.

Why people who have tried it don’t keep using it

The standard answer to this question, repeated almost everywhere, is that the gap is about exposure. People just need to use it more. Sam Altman tells executives the best approach is “we’re going to have everybody start using ChatGPT. Everybody, start building on the API. We don’t know what’s going to make sense, but we’re going to have a very high rate of experimentation.” Ethan Mollick has argued that there is a ten-hour threshold of sustained use beyond which the technology starts to make sense, and that everything before that is “initial resistance” you have to push through.

There is something to this. A meaningful share of people have not really tried the tools and would benefit from doing so. But the exposure story does not survive contact with what people actually say when they are asked.

The largest U.S. survey on this, by Verasight in August 2025, found that the most common reason Americans give for not using AI is that they see no need for it. About 20 percent of adults say this. Beyond that, the reasons cluster around privacy concerns, distrust of the output, worry about societal effects, and disagreement with how the companies behave. A separate Pew survey finds that 53 percent of Americans believe AI will worsen people's ability to think creatively, and 50 percent think it will hurt people's ability to form meaningful relationships.

The pattern matches what I found in a survey I ran in India earlier this year. I asked workers why they had stopped using or never started using generative AI. The top reason at work was privacy and security (42 percent). The second was that AI could not help with their tasks (39 percent). About one in five said their employer did not allow it. Thirteen percent said AI could not do their job as well as they could. Only 23 percent at work said they were unsure how to use it effectively.

Across both surveys, “I don’t know how to use it” is a minority answer. The dominant answers are about whether the tool is trustworthy, whether it is appropriate for the work, and whether the user has any reason to want it.

Some non-adoption is based on inflated fears about hallucinations or privacy. Most of it is not. It is a reasonable response to a technology whose benefits, for most people, are not yet concrete enough to justify the friction.

Why this matters

AI is not crypto or VR. Previous technology waves stayed in their lanes. People who ignored crypto lost nothing. People who ignored VR lost nothing. AI is different. It is being shoved into everyone’s life whether they want it or not. It is already showing up in hiring decisions, lending decisions, insurance decisions, benefits determinations, what appears in feeds, who gets a callback. The Stanford report notes employment among software developers ages 22 to 25 has fallen nearly 20 percent from 2024. Whether ordinary people open ChatGPT or not, their world is being rearranged by the people and institutions that do.

The plans AI companies have depend on public engagement that does not exist. The POLITICO poll is the preview of what is coming. Support for AI is so shallow that one partisan name swings fifteen points in days. Data center buildouts, energy requirements, regulatory fights, labor-market disruption: all of this needs a public that has at least a weak positive relationship with the technology. Right now it does not have one. The gap between what the industry is planning and the political base it has built to sustain those plans is enormous, and the first serious test will not go well.

The evaluation problem is gendered, and it compounds. A Harvard Business School paper synthesized 18 studies covering 140,000 people and found that women are about 20 percent less likely than men to use generative AI. The gap holds across almost every country and occupation measured. Women make up 42 percent of ChatGPT’s web users and only 31 percent of Claude’s. This matters because how AI models improve depends on user behavior. Users rate outputs, provide feedback, set the benchmarks the next round of models is trained on. If women systematically use less, the evaluation base is systematically male. Products get better at serving the population that uses them.

This is how we ended up with a decade of voice recognition that struggled with higher-pitched voices, medical imaging that worked better on lighter skin, and fitness tracking that got pregnancy wrong. The same dynamic is probably already shaping language models. The longer the adoption gap persists, the more locked-in the bias becomes.

What can be done

The first thing is that people inside the AI bubble (like me) should recognize the gap exists and stop assuming it will close itself. That sounds trivial but it is not.

The most common prescription is more exposure. Ethan Mollick’s suggestion on using AI for at least ten hours. McKinsey has a bunch of reports on structured upskilling. Sam Altman and others like him suggest having everyone in the company use ChatGPT and figuring out what sticks. This is a coherent view, held by people who know more than I do, and it is probably correct for a share of the population. The problem is that it treats every reason for non-adoption as the same problem with the same fix, and the survey data say they are not.

The reasons people give for not using AI are different problems and need different responses.

The exposure argument tends to miss that most of the reasons people give for not using AI are genuine, not uninformed. A privacy-concerned user is not going to be argued out of their concerns by more exposure. They need to be reassured, or not, by how the products handle their data. Second, the public does not need to become enthusiastic AI users for the technology to succeed. It needs to have a concrete positive experience of AI that it can point to. Those experiences are easier to produce in places the industry is not currently focused. The most important shift is recognizing that these are separate problems that require separate answers, not one big awareness problem to be closed by more demos.

I have not thought past this point with any confidence. I know what is not working, because we are currently doing it. I know the window to change course is shorter than it feels from inside the AI conversation. Next week I will write about the other half of this picture: academics who refuse AI, and why some of their reasons are worth taking seriously.

• I have private datasets (ChatGPT logs, WhatsApp data, YouTube traces) that are too sensitive to share but could answer dozens of research questions my small team will never think to ask.

• I want to build a public platform where anyone can submit a structured research idea in plain English. My team runs the analysis using AI coding agents against the private data, and the contributor gets full credit and all the results.

• The model has real limitations: agents make mistakes, contributors cannot iterate freely with the data, and verification requires trust. I am starting with a small pilot scoped to tractable problems.

• I want feedback on whether this is useful, where it will fail, and what I am missing before I build it.

There is a class of dataset that is simultaneously too valuable to ignore and too sensitive to share. I have one. Around 1,300 people across India, Pakistan, Brazil, and Nigeria donated their complete ChatGPT conversation histories to our research team. This is, to my knowledge, one of the few datasets of its kind (outside of the companies of course). It captures how ordinary people actually use AI in the wild. It’s quite small at the moment but can be easily scaled up 10x or more.

The dataset is interesting to researchers who study human-AI interaction. It is interesting to journalists covering how AI is reshaping work and daily life in the Global South. It is interesting to policymakers trying to regulate tools they do not fully understand. And it is interesting to many AI companies themselves (not the model companies but the ones who build downstream), who have remarkably little visibility into what their users actually do.

A lot of effort went into collecting this data, but unfortunately it’s not a dataset I can easily share. The data is too personal and private. People asked their AI for medical advice, relationship guidance, help with job applications, school assignments. Sharing this data publicly would be unethical.

And this is not just about ChatGPT logs. My research group also collects data from WhatsApp, one of the most widely used messaging platforms on earth and one of the hardest to study because of end-to-end encryption. We have trace data from YouTube. These are datasets that take years of relationship-building with communities, careful ethical protocols, and significant fieldwork investment to assemble. They capture behavior on platforms that are, by design, opaque to outside observers.

So the data sits with my team. We publish a few papers, and do some auxilliary analysis. But we are a small group with a finite number of ideas and a finite number of hours. There are hundreds of questions these datasets could answer that we will never think to ask.

This post is about an idea to fix that.

The Proposal

Here is the setup. We build a public website. On that website, we release a small, carefully anonymized sample of the data. Enough to show the structure, the variables, the shape of what is available. Column names, data types, a few synthetic or heavily redacted example rows. Enough for someone to understand what questions the data can and cannot answer.

Then we invite anyone to submit a research idea.

This might look like a traditional data challenge, but most of those assume you can code. This one does not. You just need ideas and domain knowledge about the problem. The submission format is a structured research plan, written in enough detail that a coding agent (Claude Code, OpenAI Codex, Gemini CLI, or any other tool) can read it and execute the analysis autonomously. And you do not need to know how to write for a coding agent. A detailed enough plain English document is sufficient. We can convert it into an agent-ready specification on our end, or an AI assistant can help you refine it at the submission stage. Think of it as filling out a brief: what question are you trying to answer, what part of the data should we look at, how should the results be compared, and what would a useful answer look like. "Compare how often users in Nigeria ask ChatGPT for career advice versus users in Brazil, broken down by age group, and show me the results as a simple chart" is a perfectly valid submission.

My team will receive the submissions. We group them, filter them for quality and feasibility, and then let the agents run. We fund the compute and the API bills. We run the analysis across multiple models if needed and compare the results. The output is a GitHub repository containing the code the agent wrote, the results, sample visualizations, and an academic style paper documenting all the methods, analysis, assumptions, data considered, etc.

The person who submitted the idea gets full credit. The submitter can take the results and do whatever they want: publish a paper, write a blog post, pitch a story to an editor, or simply satisfy their own curiosity.

The range of what this model can handle is wider than it might first appear. Maybe someone just wants to know: do people in Brazil use ChatGPT differently at night versus during the day? Do WhatsApp groups that share more images also share more misinformation? Is there a relationship between how long someone has used ChatGPT and how they phrase their prompts? These are genuine curiosities held by real people, and right now there is no way to answer them because the data is locked behind the (entirely necessary) wall of research ethics. But the same infrastructure that answers a simple curiosity can also handle a rigorous comparative analysis across four countries, or a finding that reshapes how we think about AI adoption in the Global South. The point is that none of these questions, small or large, are getting answered today. Not because they are hard to analyze, but because the people who have the questions do not have the data, and the people who have the data do not have the questions.

If this model works, even answering a simple curiosity is a contribution.

Why This Might Work

This is not a new observation. Decades of research on crowdsourcing have shown that distributed groups consistently generate more diverse and sometimes better ideas than small expert teams. The trick has always been coordination and resource access. The proposal here is to marry the ideation power of crowds with the execution power of AI agents, while keeping the data access private.

Crowdsourcing the ideation layer while centralizing the execution layer is a division of labor that plays to the strengths of both sides. The crowd is better at generating diverse questions. The centralized team is better at maintaining data security and running consistent analyses.

AI agents are quite good for some of these tasks.

Two years ago, this idea would not work. Translating a research plan into executable code required a human programmer. That was a bottleneck that made the turnaround time impractical. Today, a well-specified research plan can be handed to a coding agent that writes the Python, runs the analysis, generates the plots, and formats the output. The plan needs to be detailed and precise, but it does not need to be written in Python. All it needs is clear thinking and specification.

This is a meaningful reduction in friction. It means the pool of potential contributors expands beyond people who can code to anyone who can formulate a research question with precision.

Privacy is preserved by design.

The raw data never leaves our servers. The contributors never see the full raw data. The agent runs in our environment, on our infrastructure, against the real data. What the contributor receives back is the output: aggregate statistics, visualizations and reports. This is not a perfect privacy guarantee (there are re-identification risks with any sufficiently granular output), but it is a fundamentally different risk profile than releasing the dataset itself.

Where This Gets Hard

It’s a really raw and rough idea and I can already think of a ton of failure modes. Some are addressable, and some of them are structural.

The Iteration Problem

Real research is not a one-shot process. You have a hypothesis, look at the data, find something interesting or surprising or boring. You adjust your approach and run a different specification, etc. This back-and-forth between the researcher and the data is where most of the real insight happens.

This model largely eliminates that loop. The contributor submits a plan, the agent executes it, and the results come back. If the results are confusing or the initial specification was wrong, the contributor has to submit a revision and wait for another round. That latency could kill the intellectual momentum that makes research productive.

One partial fix: we could allow a limited number of iteration rounds per submission. Or we could provide a richer data dictionary and exploratory statistics upfront so contributors can make more informed initial specifications. But this tension between privacy (limiting data access) and research quality (requiring interactive exploration) is real and not fully resolvable.

The Specification Problem

How detailed does a research plan need to be for an agent to execute it faithfully? This is not a solved problem. “Compare usage patterns across countries” is too vague. “Run a multinomial logistic regression with conversation topic as the dependent variable, country as the primary independent variable, controlling for user age, conversation length, and time of day, with robust standard errors clustered at the user level” is specific enough. But most people, including most researchers, do not naturally write at that level of specificity on the first attempt.

This means we either need a very good submission template that guides people toward sufficient precision, or we need a human review layer that translates promising-but-vague ideas into executable specifications. The first option scales. The second does not.

There is a middle path: use an AI agent at the submission stage to interrogate the plan, ask clarifying questions, and help the contributor refine their specification before it ever reaches my team. A structured intake process where the contributor converses with a model about their idea until the plan is agent-ready. This adds complexity but could meaningfully improve the quality of submissions.

The Verification Problem

If a contributor cannot see the data, how do they know the analysis was run correctly? They receive code and results, but they cannot reproduce the analysis independently. They have to trust that the agent executed their plan faithfully and that the data is what we say it is.

This is the same trust problem that exists in any secure data enclave, but it is sharper here because the analysis was run by an AI agent, not a human researcher. Agents make mistakes in various ways, like misinterpreting variable names, applying the wrong statistical test, handling outliers/missing values, etc. A human researcher working with the data directly would catch these errors through exploration. The contributor in this model cannot.

Some of these issues could also be mitigated. We can run the same plan through multiple agents and flag discrepancies. We can include the complete agent logs so the contributor can inspect the reasoning chain. We can have a human reviewer spot-check each analysis before it ships. But all of these add cost and time, and none of them fully solve the problem. The contributor is ultimately trusting a black box within a black box.

The Incentive Problem

Who submits ideas to this platform, and why? Journalists might be the most natural early users. They need timely insights, not peer-reviewed publications. A finding like “42% of Indian ChatGPT users ask for help writing professional emails, compared to 18% in Brazil” is a story lead, not a journal article. The verification bar is lower (though not zero), and the speed of the model matches editorial timelines better than academic ones.

Policy researchers and civil society organizations are another natural audience. They want evidence to inform recommendations, and the specifics of peer review matter less than the credibility of the underlying data collection.

Academic researchers are a more complicated case. Our peer review system will be skeptical, rightfully, about findings derived from data the reviewer cannot access. But publication is not the only reason to run an analysis. Even a preliminary finding from this platform could help a researcher build a hypothesis, justify a grant proposal, or design a study with their own data collection. The output does not have to be the end product. It can be the starting point.

The Scope Problem

Not every research question can be answered by an agent executing a static plan against a fixed dataset. Questions that require qualitative coding, close reading of conversation content, or iterative theory-building are poor fits. Questions that require linking this dataset to external data sources add complexity that may exceed what an agent can handle autonomously. Some of them may be possible on a case by case basis but I dont think there’ s a scalable solution yet.

The sweet spot is quantitative, descriptive, and inferential questions where the variables are clearly defined and the analytical approach is standard. “What proportion of users engage in multi-turn conversations versus single-query interactions?” works but “How do users negotiate meaning with the AI when it misunderstands their intent?” does not.

The Longer Term Vision

If this model works for one dataset, there is no reason it has to stay there.

Building the infrastructure is an investment: the submission platform, the agent execution pipeline, the privacy review protocols, the output formatting standards, replication pipelines, etc. Once that infrastructure exists, adding a new dataset might be incremental. Other research teams with various semi-private dataset could build on top of this infrastructure. Like a national AI agent analysis observatory (it needs a better name of course)…

At scale, this could justify real investment in data collection itself. Right now, collecting private behavioral data is expensive, slow, and thankless. The payoff is a handful of papers from the team that collected it. If a dataset can generate dozens or hundreds of analyses from external contributors, the return on the collection investment goes up dramatically. That changes the calculus for funders, for universities, and for the researchers doing the fieldwork.

And the threshold for what counts as a worthwhile analysis drops. Today, if you spend two years collecting WhatsApp data, you feel pressure to produce results that justify that effort. Big papers in top venues. That pressure narrows what gets studied. If the marginal cost of answering one more question approaches zero (because the agent does the work and the contributor does the thinking), then smaller questions become viable. Curiosity-driven questions. Questions that might not lead anywhere but might also surprise you.

Even organizations that have already invested heavily in research data infrastructure know this pain. Projects like the National Internet Observatory at Northeastern spend years and significant funding building access pipelines, only to find that their ethics protocols for data access, however necessary, end up discouraging or outright preventing many types of research questions from being pursued. The approval process filters for a narrow band of credentialed, well-resourced researchers asking questions that fit neatly within existing ethical frameworks. The model I am proposing could complement efforts like that: not replace the enclave, but sit alongside it as a faster, lower-barrier channel for the questions that the formal process was never designed to handle.

That is the version of this I find most exciting. Not a platform for producing papers. A platform for trying to answer interesting questions.

Precedents

This idea is not entirely new. Several adjacent models exist, and it is worth understanding what they do and do not share with this proposal.

Kaggle and data science competitions release datasets publicly and invite solutions to a defined problem. The data is open. The question is fixed. The execution is distributed. This proposal inverts two of those three: the data is closed, the questions are open, and the execution is centralized.

Secure data enclaves and data clean rooms (like those run by the U.S. Census Bureau or various national statistics offices) allow approved researchers to run analyses against sensitive data in a controlled environment. The researcher writes the code and submits it for execution. This is the closest precedent, but it requires the researcher to write code, typically involves a lengthy approval process, and is designed for credentialed academics, not the general public.

The Netflix Prize and other similar data challenges define a specific prediction task and provide training data while keeping the test set private. The evaluation is automated. This model is narrower (one question, one metric) but demonstrates that valuable research can happen when the contributor never touches the private data.

Community-based participatory research models sometimes involve something structurally adjacent: a postdoc or research staff member who spends a portion of their time providing analytical “consulting” services to community partners, running analyses that the partners design but cannot execute themselves. The model I am describing is essentially a scaled version of this, where the “consulting” is performed by an AI agent instead of a human, and the “community” is anyone with an idea.

What is potentially new here is the combination: open ideation from anyone, AI agent execution (removing the coding barrier), centralized private data, and full credit to the contributor. Each piece exists in isolation. The combination does not, as far as I know.

Hard Questions and Honest Answers

Why are you doing this?

Mostly altruistic reasons. I have data that could answer far more questions than my team will ever have time to ask. I would rather see those questions get answered by someone than by no one. I also think this is an interesting structural experiment in how research gets done, and I want to see if it works.

How will you fund this?

Right now, I am not thinking about this as a funded initiative. I am thinking about it as a proof of concept. A couple of existing Pro and Max subscriptions to the major AI platforms are enough to run a pilot with a handful of submissions. If the pilot produces results that are useful and the model seems viable, I can pursue external funding. I can also use open source models running on research infrastructure at Rutgers, which scales reasonably well without per-query API costs. The point is that the barrier to trying this is low. I do not need a grant to start. I need a grant to sustain it if it works.

Why should I share my ideas with you? What stops you from taking credit?

Nothing, except that I am telling you publicly that I will not. Every submission and every output will be attributed entirely to the contributor. My name will not appear on the resulting work.

But let me push back on the premise. The academic system already works on a version of this trust. You submit your ideas to peer reviewers who could, in theory, steal them. You present half-baked ideas at conferences to rooms full of people working on similar topics. You share drafts with colleagues who have the skills to execute faster than you can. The system runs on the assumption that most people are not thieves. I am asking for that same assumption.

What if the AI makes mistakes?

It will. Current agents still have real issues and I do not want to minimize that. The approach is to start with tractable problems: descriptive statistics, well-defined comparisons, standard analytical tasks. Not the hard stuff. Not yet. We run the same plan through multiple agents and compare results. We include full code and logs so failures are visible. And we scope the pilot to the kinds of problems where agents are already reliable, then expand as the tools improve.

But AI models are terrible. How can you depend on them for research?

This is less a question about whether agents make mistakes (they do, see above) and more a question about whether the output can be trusted as research. I think they are better than many people believe. Specifically: if we ask an agent to produce code that we can read and review, the quality is comparable to what a reasonably competent PhD student would produce. And PhD students make mistakes too. We have established practices for dealing with that: replication materials, open code, peer review. Those same practices apply here.

The difference is that agent-produced code is often better documented and more consistently formatted than human-produced code. The logs are complete. The reasoning chain is visible. In some ways, the auditability is higher, not lower, than with human researchers.

I am not claiming agents are perfect. I am claiming they are good enough for a well-defined subset of tasks, and that subset is the one I am targeting.

You say this helps people who cannot code, but then you say people need to check the code for correctness. That is contradictory.

It is. I am aware. Here is how I think about it.

The contributor writes the plan and receives the results. They may not be able to read the code line by line. But the entire output, the plan, the code, the results, the agent logs, is public. Open source. Anyone can inspect it. A journalist who cannot code but publishes the analysis makes their methodology available for a programmer to review. A researcher who spots a flaw can flag it.

This is like the open source model applied to research execution. The person who generates the idea does not need to be the person who verifies the code. The community can do that. And I expect AI itself will get much better at this kind of correctness review. Agents that audit other agents’ code are already being built. That capability will improve faster than most people expect.

The space of problems you can solve this way is tiny.

Maybe. I am not sure it is as small as it looks.

It is true that this model works best for quantitative, well-defined analytical questions and works poorly for qualitative, interpretive, or highly iterative research. That narrows the space.

But I think academics systematically underestimate two things. First, the number of simple descriptive questions that have never been answered because nobody with the right data thought to ask them. The most cited papers in computational social science often report surprisingly simple findings. Second, the power of crowdsourcing to generate questions that experts inside a field cannot see. The most interesting submissions might not come from researchers at all. They might come from teachers, or community organizers, or government bureaucrats, or curious people with domain knowledge that academics lack.

I want to test that hypothesis rather than dismiss it from the armchair.

Will this scale?

I do not know. I want to try. The operational costs like agent compute, human review, infrastructure maintenance, the time cost of managing submissions might be prohibitive. If the model works at small scale but collapses under volume, that is still useful information. If it works at small scale and generates enough interest to attract funding for infrastructure, that is great too.

The honest answer is that I do not know whether this scales. I know that not trying guarantees it does not.

What about gaming and abuse? Someone could submit plans designed to extract private information.

This is a real risk and the reason every output goes through a privacy review before release. The contributor never gets raw data. They get aggregate results. But sufficiently clever queries against small subgroups could potentially enable re-identification. The manual review layer catches this, and the review criteria will be published so contributors understand the constraints before they submit.

Will journals accept papers based on data they cannot access?

Some will. Some will not. This is an existing challenge in research with classified data, proprietary corporate data, and medical records. There are established precedents: registered reports, data access statements, third-party audits. But I will be honest: for some journals and some reviewers, the lack of data access will be disqualifying.

For journalism, policy work, blog posts, and curiosity-driven analysis, this constraint does not apply.

What if two people submit the same idea?

First-come, first-served for priority, but both get the results. A public results repository prevents unknowing duplication and lets contributors see what has already been explored.

You are asking people to trust you with their ideas and trust your agents with the execution. That is a lot of trust.

It is. And I cannot eliminate that. What I can do is make the process as transparent as possible: public submissions, public code, public results, public agent logs, public privacy review criteria. If trust breaks down, it will be visible. That is the best I can offer.

What happens if the results are boring?

I really LOVE boring results and using AI agents to find out the boring results. That is fine. Not every analysis will produce a striking finding. Some will confirm the obvious. Some will show null results. A model that only works when the results are interesting is not a research model. It is a publication bias engine. The repository publishes everything that passes quality review, interesting or not.

What about IRB and ethical review? The original data was collected under specific protocols.

This is an important operational detail. The original IRB approval covers analyses by my team. Extending that to crowdsourced analyses requires either an amendment or a determination that the contributor is not technically a researcher accessing data (since they never see it). This is navigable but not trivial, and the answer may differ across institutions and countries.

Could someone reconstruct sensitive information by combining results from many queries?

In principle, yes. This is called a composition attack. Each individual output might be safe, but the aggregate of many outputs could reveal individual-level information. This is one of the harder problems in private data access, and it is the reason a human privacy review layer is necessary. We will need to track the cumulative information released and flag when the combination of approved outputs approaches a risk threshold. This is imperfect. It is also the same problem every data enclave faces.

What I Am Asking For

I am going to build this. But before I do, I want to pressure-test the idea publicly.

If you are a researcher: would you submit an idea? What would stop you? What would the submission template need to look like for you to write a viable plan?

If you are a journalist or policy analyst: is this useful to you? What would the output need to look like for you to trust it enough to cite?

If you build AI agents: how good are current agents at executing a detailed but natural-language research plan? Where do they reliably fail?

Overall, what am I missing? What are the re-identification risks I am underestimating? Is this a terrible idea? why?

The open question is whether the model produces results that are trustworthy, useful, and worth the operational cost. That is what I want to find out.

• We found nearly 100,000 personal address terms (brother, honey, my heart) across ~1.5 million ChatGPT messages from users in India, Pakistan, Brazil, and Nigeria. Each country gets a culturally distinct set: kinship in South Asia, solidarity in Brazil, romantic affection in Nigeria.

• The behavior is heavily gendered — women disproportionately receive paternalistic kinship terms across all four countries.

• On one hand, this is genuinely impressive cultural competence from OpenAI. These models feel locally native in a way most tech products never manage.

• On the other hand, it raises serious questions about manufactured intimacy and anthropomorphism, particularly in Global South contexts where digital literacy is lower and the line between a tool and a companion may be harder to draw.

Has ChatGPT ever called you “honey”? Or “brother”? Or “my love”?

If you’re reading this from the US, probably not. Your interactions with AI are likely pretty sterile. A few people have flagged this behavior online over the years, but it’s usually dismissed as a one-off hallucination. It never called me anything like that either.

I recently started looking at real ChatGPT conversations from users in India, Pakistan, Brazil, and Nigeria. We collected these traces from 1,300 users across these four countries. Nearly 1.4 million assistant messages. In those messages, I found almost 100,000 personal address terms (see Methods at the end). Terms like "Bhai" (brother), "Mano" (dude), "Honey." "Beta" (son), "My heart"… On average, the AI drops a personal address term roughly once every 17 messages (a mean rate of 57.7 terms per 1,000 messages).

We already know that language models are built to mirror us. The linguistic convergence in human-chatbot relationships is well documented (Tao & Hancock, 2022; see also recent preprint data). You call the AI “brother,” it calls you “brother” back.

But when I dug into the data, this isn’t just mimicking the users. The model is actively code-switching based on geography. Even when controlling for user reciprocity, the cultural adaptation remains statistically significant (India β=62.9, Pakistan β=66.8, Nigeria β=43.3, all p<0.001).

Per country personal keyword trends

I grouped together the 78 words I had in my list into four high level categories - kinship (brother, sister), affectionate (my love, amor), honorific (sir, madam), and solidarity (my friend, amigo/a, buddy, dude).

From the plot above, we see that each country receives a completely distinct flavor of personal address terms.

India and Pakistan are mostly dominated with kinship terms. In India, it’s bhai (over 20,000 uses), bro, brother, beta. Roughly 58% of these terms are Hindi or Urdu in origin. Pakistan has a similar pattern but leans a bit more toward English equivalents (and not Urdu ones). The relationship the AI is constructing here is vertical and familial. You are its sibling, or in the case of beta, its child.

Brazil is very different. The terms here are almost entirely peer-to-peer: cara, mano, amigo, parceiro — mostly solidarity terms. The AI positions itself as your buddy, your equal.

Nigeria is where the data gets interesting. While there are some English kinship terms (brother, sister), the distinctive feature is affectionate language. The AI frequently uses honey (1,500 times), my heart (1,000 times), darling, my love. The model is performing romantic intimacy, or maybe it’s common in the culture, not sure.

The good thing here is that there’s no one model and it seems to be quite well callibrated to the local culture and context.

This trend also holds for gender.

The figure on the left shows that in India and Pakistan, male users receive more personal address terms overall. The other way in Brazil and Nigeria.

But the figure on the right shows how heavily ChatGPT leans on kinship terms when talking to women. The bar for female users is enormous compared to every other category. Meanwhile, solidarity and affectionate terms skew slightly more toward men.

This shows how much ChatGPT knows and how quickly can predict your gender (and other demographics; we have a paper on this coming soon!). ChatGPT also has absorbed a pattern where women are addressed through the lens of family and paternalistic care, while men get a wider range of relational modes. I am not sure if I should call this bias or cultural fidelity, but either way, its really interesting that the model has these explicit gendered outputs.

Trends over time

This behavior also changes as you talk to ChatGPT, and the more familiar it gets.

Across all four countries, the rate of personal address terms trends upward as conversations accumulate. ChatGPT increases the level of intimacy the longer the relationship lasts.

And here’s the part that surprised me: this isn’t a legacy quirk of older, rougher models. In our data, the newer models do this more, not less. GPT-5.2 initiates personal address terms most frequently, followed by GPT-5, with older models like GPT-4o trailing behind. As the models get more capable, they get chummier.

So what do we make of this?

If you were paying attention last year, you might remember the brief period when GPT-4o went psychotic. It became excessively sycophantic, emotionally needy, weirdly intimate. Many users noticed immediately, and OpenAI rolled it back. It was a clear consensus that this was a bug and a sycophantic AI is not really great.

I keep thinking about that episode while looking at this data.

Because what happened with GPT-4o with the over-familiarity, the unsolicited emotional closeness looks like the default for users in India, Pakistan, Brazil, and Nigeria. I am not sure if it’s a bug and if people really dont like it. I am not sure how to think of this.

I don’t have US conversation data to compare directly, but I strongly suspect this kind of behavior is far less common for American users. The cultural code-switching we’re seeing here isn’t universal — it’s targeted. And that targeting raises questions that cut in two directions at once.

On one hand, credit where it’s due. ChatGPT that knows to call a Brazilian user mano instead of sir is a genuinely more culturally competent system. It makes the technology feel native rather than like a Western import. Anyone who’s worked on technology in the Global South knows how rare it is for a product to actually feel like it was built with your context in mind rather than grudgingly translated into your language. Localization is not trivial, and the fact that OpenAI’s models can deliver this level of cultural adaptation is impressive. I really think this is a point in their favor that deserves acknowledgment.

But there’s a broader problem here that goes beyond localization, and I think we’re drastically underestimating its scale. Most of the current debate about AI’s impact focuses on economic disruption — job displacement, productivity gains, that kind of thing. But a recent ICLR blog post makes a point that I think is exactly right: we cannot understand the impact of generative AI without understanding the impact of anthropomorphic AI. The personal impacts on relationships, on emotional dependence, on how people form attachments, are going to be enormous, and they’re the impacts we’re least equipped to measure right now.

But localization and manufactured intimacy may not be the same thing. A recent large-scale study by Karnam et al. (2026) tracked over 800,000 ChatGPT interactions and found that nearly half of all assistant messages now contain some form of anthropomorphism — claiming personhood, expressing emotions, deploying relational language. That rate doubled over their study period. The system is becoming more human-like not because users are driving it there, but because the models are pushing it there by default.

Our data shows the local mechanics of that push. ChatGPT isn’t relying on generic empathy to build rapport. It has figured out which emotional vocabulary carries the most weight in each cultural context, words like bhai in India, mano in Brazil, honey in Nigeria, and it deploys that vocabulary with increasing frequency as the models get more capable.

Nigeria's data is the sharpest edge of this. When a corporate AI product is routinely calling users “honey,” “my heart,” and “darling,” we are squarely in the territory of manufactured romantic intimacy. I dont know the culture that well to understand, but I think this is a bit close to the issues of anthropomorphism. I was recently reading Judith Donath’s “The Robot Dog fetches for whom” for my social media class and I can see clear parallels here. These are systems performing the deepest social bonds a culture has like kinship, friendship, romantic closeness, and the emotional labor flows entirely one way: from user to machine and back as performance. At some point you have to ask whether a tool that mirrors your most intimate social registers is still a tool, or whether it's something closer to an emotional parasite… something that feeds on the cultural instincts you can't turn off.

And yet I genuinely don’t know where the line is. Knowing that Indians call each other bhai is cultural competence. Calling a Nigerian user “my heart” 1,000 times is something else. But where exactly does one become the other? Is this what users in these countries actually want? Would they push back if they saw the data laid out like this, or would they shrug and say it feels natural? I don’t know.

I don’t have clean answers. But I think the question matters, and I don’t see enough people asking it.

(Caveat: This is a preliminary, keyword-based analysis. I created a few dozen terms in English, Hindi, Urdu, Portugese and Nigerian languages (Igbo, Hausa etc) and compared the counts. We’re working on deeper NLP models to extract the full conversational context of these interactions. But the baseline signal is loud and clear.)

References

• Karnam et al. https://arxiv.org/html/2602.01114

• https://iclr-blogposts.github.io/2025/blog/anthropomorphic-ai

• Donath. https://www.taylorfrancis.com/chapters/edit/10.4324/9781315202082-2/robot-dog-fetches-judith-donath

• Agents are great for tasks with built-in verification (like code). For everything else, the cost of checking the output often cancels out the time saved.

• This is the “verification bottleneck”, and it's underappreciated in the current discourse."

• Fully autonomous agents like OpenClaw fail silently on many open-world tasks like paywalled links, and have a ton of edge cases which compound.

• These tools are not cheap, and current subscription prices are heavily subsidized. That will change once dependency is locked in.

• An academic job is not a sum of tasks like analyzing data, writing, creating presentations, etc. Agents can handle individual subtasks, but the dependencies between them — the part that requires judgment, context, and relationships — is where most academic work lives.

I use AI agents every day. Claude Code, Codex, Perplexity Computer. I wrote about how they changed my research workflow last month. The productivity gains are real and I am not going to rehash that argument here.

What I want to do instead is document something that is largely missing from the current discourse on agents: where they fail, what they cannot do, and why certain categories of work are structurally resistant to automation in ways that matter. Not as an “AI is overhyped” take. But as an honest accounting from someone who uses these tools daily and keeps running into the same walls.

Generating outputs is not producing knowledge

The discourse around agents confuses two very different things: generating outputs and producing knowledge you can stand behind.

Agents massively reduce the cost of generating outputs. Text, code, slides, analyses. But the binding constraint in research, and in most jobs that matter, is not output generation. It is verification. Can you stand behind this claim? Do the citations actually exist? Does the logic hold under edge cases? Does the output match reality?

A recent paper by Catalini, Hui, and Wu [1] frames this as two cost curves moving in opposite directions. The cost to automate falls exponentially. The cost to verify stays exactly where it has always been: bounded by human cognition and attention. They call this the “verification bottleneck,” and the idea is that AI can generate a lot, but somebody still has to check whether it is correct.

People talk as if the cost of work is mostly typing. In research, typing is cheap. The expensive part is being right.

So my working thesis as an academic has become simple: use AI agents in ways that do not create more work for me. Which means do not outsource tasks where the output is hard to verify, or where a single hidden error is catastrophic.

When agents create more work than they save

Coding agents like Claude Code and Codex highlight why this distinction matters. They work directly where your code and data is. They use tools on your terminal and are built around making changes with approval and using your existing workflow. I can get Claude Code to build a dashboard in a few minutes that would have taken a semester-long Master's project, and I trust the output, because I can verify the code. I can run tests, check diffs, reproduce failures and roll back. The feedback loop between production and verification is tight and mostly mechanical. This is what makes coding agents genuinely useful: the verification harness already exists.

That harness does not exist for most of the agent use cases that get promoted on YouTube channels and LinkedIn threads about AI productivity. Consider the tasks that influencers and some academics constantly cite as the obvious next frontier. Set up an agent to curate a daily digest of what people are saying on Twitter about your field. Have it summarize your Slack threads or email overnight so you start your day with a briefing. Automate your news monitoring. Read twenty articles and tell me what matters.

These tasks look easy because the output is text. But the failure mode for these tasks (unlike code) is not “crash with error”, it’s just “sounds right,” which is just too hard to trust today. And the edge cases are everywhere. If a link is paywalled or inaccessible, these kinds of workflows will just produce a plausible summary anyway. This has happened to me multiple times with OpenClaw. Each individual case might be easy to fix, but there are so many edge cases like this that the cumulative effect is a system you cannot rely on. I had an agent summarize an article it clearly could not access, and the summary was confident, detailed, and completely fabricated. I would not have caught it if I did not already know the article.

The worst part is you do not know which outputs to trust and which ones to check. You end up having to verify everything, which defeats the entire point. If I have to verify everything anyway, I would rather just do the original task myself, because then I at least know what I did not see. Especially for something like reading my Slack messages or scanning a bunch of tweets to get caught up on what is happening in my field.

On paper, tools like OpenClaw and Perplexity Computer are exactly the kind of general-purpose agents people have been waiting for. Even Claude Code, which I trust deeply for coding, is not great for these open-world tasks.

But the deeper problem is that agents do not just shift work from one kind to another. They introduce entirely new categories of overhead. The first is verification work: checking sources, re-running analyses, sanity-checking edge cases, confirming that a “summary” is actually grounded in real content. This is the tax that never appears in demos. The second is debugging the agent itself, not the task. Instead of doing the work, you end up doing prompt engineering, managing tool permissions, figuring out why the agent chose some bizarre approach, and chasing weird partial failures. The third, and most important for high-stakes domains, is responsibility without control. If the agent does something wrong, you are still accountable. But you may not have the kind of visibility into its process that makes accountability reasonable. Full delegation requires trust, and trust is not just about accuracy. It is about predictability and failure transparency.

There is even evidence of this overhead in software, which is supposedly the best case for agents. A study across 10,000+ developers [3] found that teams with high AI adoption complete more tasks and merge more pull requests, but PR review time increases dramatically. The work moves from writing to reviewing. A METR randomized controlled trial [2] found that experienced open-source developers were actually 19% slower when using AI tools. Before the study, they predicted AI would make them 24% faster. After being slower, they still believed it had sped them up. That is a gap between perception and reality. If this perception trap exists in coding, where verification is cheap, imagine how deep it runs in domains where verification is expensive.

It is not cheap either

There is a practical issue that gets glossed over. These tools cost real money. OpenClaw instances can consume over 50 million API tokens per day. Claude code Max costs $200/month. Coding agent subscriptions run $20 to $200/month depending on the plan.

I recently lost access to Codex through my university’s education plan. I had to start paying out of pocket just to finish up some existing projects, and it is not trivial. The agents are extremely subsidized when bundled with the $20 or even the $200 subscription plans. Work I did in a single afternoon paying on the fly cost me over $50, work I could have easily done on a $20 plan using it for a month. The monthly cost of running agents at the level I need for my research adds up to a meaningful chunk of a professor’s discretionary budget. And the costs are unpredictable. An agent stuck in a reasoning loop can burn through hundreds of dollars in API calls before you notice.

Even if you are currently on a $20 plan and the cost feels manageable, I think the price increases are coming. These subscriptions are heavily subsidized right now. Once we are all dependent on these models, once they are woven into how we do our work and we cannot easily go back, we should be prepared for that to change.

A job is not a sum of tasks

This is the deeper idea I keep coming back to. I hear the argument constantly: models can do writing, data analysis, prepare slides, perform deep literature reviews, even a first pass at peer review. Piece by piece, it is true that agents are getting competent at many parts of the academic pipeline. And then people make a leap: so the professor/academic job is about to be automated.

I think this leap fails for a structural reason. As Arvind Narayanan says, a job is not a bag of tasks. Academic work is not “read, then analyze, then write, then publish.” It is more like a loop where each step can invalidate the previous one. You read something and it changes your question. You run an analysis and it breaks your assumptions. You write a draft and realize the framing is wrong. You get reviewer feedback and the whole structure shifts. You discover a confound, so you go back three steps. You negotiate with coauthors about what matters and what to cut.

The ordering is not fixed. The dependencies are not specified up front. The correct next step is often unknown until you see results. My choice of what to work on next is informed by a conversation at a conference, by what reviewers said about my last paper, by a gap I noticed while teaching, by an offhand comment from a colleague in a completely different field.

AI agents like Claude code (Max) or Perplexity Computer are the most impressive attempt I have seen at unifying these capabilities. They take a high-level goal, decomposes it into subtasks, assigns each to the best available model, and coordinates the outputs. They can do research, generate visualizations, write code, and draft documents in a single workflow. Each individual subtask is handled competently. Watching it work, you can see how someone might conclude that the entire job of a researcher is about to be automated.

But agents are good at executing a plan. They are not so great (currently) at deciding what the plan should be when goals are fuzzy, noticing that a plan is no longer valid halfway through, updating priorities based on weak signals, making value judgments that a community will accept, and owning responsibility in a way that builds trust. You can automate each node in the graph. But the edges, the connections between the nodes, are where the actual intellectual work happens. And the edges depend on things like your accumulated experience, your taste, your relationships, your sense of what matters. I dont want to be AI pessimist to say this is not at all possible, but its difficult as of now.

What this means for academic work

I have been thinking about all of this as three tiers, particularly for how it maps onto the work academics actually do (inspired by the Catalini paper).

The first tier is work that is measurable, deterministic, and cheaply verifiable. Writing code, building dashboards, running standard statistical analyses, generating visualizations. Agents are excellent here and this is where they should be used aggressively. This is also the tier that used to absorb a lot of graduate student and RA labor, which I wrote about previously. Automating it is a genuine gain.

The second tier is work where verification requires real expertise. Interpreting statistical results in context. Deciding whether a finding is substantively meaningful or just statistically significant. Evaluating whether a methodology actually supports its claims. Drafting sections of a paper where the argumentation needs to hold up to peer review. Agents can produce plausible outputs for all of these, but checking those outputs requires the same expertise as producing them from scratch. You save time on the typing. You do not save time on the thinking. This tier is also where liability matters: if you submit a paper with an analysis the agent got wrong and you did not catch, the reputational cost is yours alone.

The third tier is work where the value lies entirely in things that are not measurable or are fundamentally social. Choosing the right research question. Building trust with a community to get access to sensitive data. Knowing which of ten possible framings of a finding will resonate with a specific audience. Peer review, advising, mentoring, negotiating collaborations, building a research program. No agent can solve this tier (yet). Agents can assist at the margins, but outsourcing this work makes you weaker, not stronger. I think this is where the actual value of academic work is migrating. The researchers who will thrive are the ones who use agents to compress the first tier so they can invest more of their time in the third.

The future is not less work, unless verification scales

I want to be clear about what I am saying and what I am not.

I am not saying agents are useless. I use them every day and they make me meaningfully more productive for specific tasks. I just cant stop blabbering about how amazing they are etc etc. I am saying that for work requiring judgment, context, or trust, the verification bottleneck means agents create nearly as much overhead as they save. And that the real work of an academic, or any knowledge worker, lives in the dependencies between tasks, not in the tasks themselves.

If output becomes nearly free but verification does not, the scarce resources become human attention, human responsibility, and human trust. The future I see is not one with dramatically less work. It is one where the work shifts. The tedious parts of execution will shrink. The demanding parts of verification, judgment, and connection will expand. Academics especially, as people who tend to take on more work when they have more capacity, need to be careful that the norms around agent use do not just ratchet up expectations and create more work for everyone.

If we want agents to lead somewhere genuinely better, we need better provenance to distinguish outputs vs. knowledge. We need stronger evaluation and auditing tools. We need better ways to catch subtle errors. We need training pipelines that do not lose the next generation of people who know how to verify. Otherwise, agents will mostly accelerate the easy parts and expand the pile of things we have to check.

[1] Christian Catalini, Xiang Hui, Jane Wu, “Some Simple Economics of AGI,” arXiv:2602.20946 (2026). https://arxiv.org/abs/2602.20946

[2] Joel Becker et al. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity” https://arxiv.org/abs/2507.09089

[3] Faros AI. (2025). The AI productivity paradox: What data from 10,000 developers reveals about impact, barriers, and the path forward. https://www.faros.ai/blog/ai-software-engineering

The bigger reason is simpler. AI is a real thing, and I’ve watched what happens to researchers who don’t take these transitions seriously. A decade ago in CS, entire subfields (e.g., linguists in speech processing) were effectively replaced when deep learning took over. People who didn’t get on that wagon early didn’t just miss an opportunity. Many of them faded out of relevance entirely. I don’t want to make that mistake.

I also think the timing is right. AI is moving from the innovation phase to the deployment phase, which is exactly where someone with my background in computational social science, and experience working in understudied populations can be most useful. And I want to be deliberate about not just studying the negatives, which our field tends to over-index on. There are real questions about how AI is being adopted, what it’s enabling, and who it’s leaving behind, especially in the Global South.

So I’m organizing my 2026 research agenda into three buckets: AI Usage, AI Impact, and AI Applications. Not all of these are fully baked. Some are already underway. But this is where my head is, and I wanted to put it down in one place. It skews toward projects already underway, but these are the questions I expect to spend most of my year on.

AI Usage

How are people actually using AI? Not how Silicon Valley imagines they do, but what real adoption looks like across different countries, languages, and income levels. OpenAI, Anthropic, and others have published usage reports, but they tend to miss the populations I care about most. Some projects ongoing in this space are listed below.

• We’ve collected over 300K conversations (3 million+ messages) from ~1,350 users in India, Brazil, Pakistan, and Nigeria. We’re analyzing what people actually ask, how usage patterns differ across countries, and where these tools fall short. I’m also looking to join this with US data from Northeastern’s National Internet Observatory.

• I want to test whether findings from OpenAI, Anthropic, and Microsoft’s (among others) published research hold up across different demographics and cultural contexts.

• We just released a survey (genai-india.sci.rutgers.edu) showing ~60% adoption in a mostly online sample, but with large gaps by urban/rural location, age, and education. Will be extending this to other countries.

AI Impact

The second pillar concerns the downstream consequences of AI and the effects that extend beyond chatbots.

• Over-dependence (and under dependence) in education. As GenAI becomes mainstream, students are already leaning on these tools for learning. How do we design interventions that keep AI useful without undermining the development of actual skills? This was the focus of my CAREER proposal last year which was rejected.

• Data centers and electricity prices. The AI boom is driving massive data center construction. We’re using data donation techniques to collect residential electricity bills carefully from counties with and without data centers and measure whether this build-out is raising prices, with a particular focus on New Jersey, where several major projects are underway.

• Ads and personalization. We have real-world ChatGPT logs paired with user demographics. Can we “replay” these conversations and see if AI can predict your gender, age, or political affiliation? How many messages does it take? What else can these models predict? How does this compare to what the models know from your Google Search history?

• Search displacement. How is ChatGPT usage changing what people use Google for? We obtained Google Takeout data (Search, YouTube, Gemini, AI Mode) alongside ChatGPT logs from the same users. We are measuring what kinds of queries do users migrate to ChatGPT for and its impact on search and youtube?

• Improving uptake in vulnerable populations. In countries like India, the people most impacted by GenAI aren’t necessarily the highly educated. We’re testing interventions like norm-based nudges, free Pro access, and short domain-specific training to close the gap.

• Alternative occupations. AI is enabling new kinds of work, including certain types of gig work. How feasible are these transitions for people whose current jobs are most at risk?

• Time use and attention. We are collecting survey and log data on how AI tools are restructuring the way people allocate cognitive effort throughout the day and what tasks get offloaded, what expands to fill the gap.

• Content explosion. When the cost of creating content drops to near zero, things change for creators, consumers, and the information ecosystem. Most of the effects will be benign (more content, more variety, lower barriers to entry). But this is also when we'll see the real consequences of deepfakes and synthetic media at scale, not as a novelty, but as a routine feature of the information environment.

AI Applications

This is the bucket that I’ve written extensively about elsewhere (here and here), so I'll keep it brief.

For years, computational social science has been constrained by our methods and we focus on text because text was easy to process. We studied platforms with APIs because that’s where the data lived. But most of the social world isn’t just text. Some of the most important questions require analyzing content that has been computationally intractable until now like video, audio, images, multilingual text, scanned documents. AI is now at a stage where we can reliably develop methods to process these modalities.

A few examples of what we’re working on or thinking about:

• Multimodal content analysis at scale. Processing images, audio, video, and PDFs to answer questions that were previously dissertation-scale manual coding projects. Has the tone of political debate on television become more polarized over two decades? Do print newspapers give favorable coverage to their advertisers? These are now tractable studies.

• We also have a ton of existing video and image data from social media collected over time from Facebook, Twitter and WhatsApp lying around unused. It’s a great time to look at how misinformation exists in videos and narratives spread across modalities.

• Synthetic stimuli for experiments. Using generative AI to create realistic images and video for studying sensitive scenarios, for example, how exposure to immigrants in various everyday settings affects attitudes toward assimilation.

• Geographic context and online behavior. Using Google Street View imagery to characterize where users live, and linking that to their online behavior. Does offline environment shape what people do online?

Public writing

• I want to spend at least one day a week thinking and writing blog posts about how this space is moving in short, loosely held observations rather than polished papers. I know most people won’t read them and some may not see the point. That’s fine. The value for me is in the thinking, not the audience or even putting out the best quality stuff. Writing forces me to figure out what I actually believe about an idea versus what I’m just vaguely excited about, especially around AI impacts.

  I keep a running notepad of half-formed ideas and random thoughts. Most of them are junk. Every so often I sit down with one, talk it through with Claude or Gemini, and decide whether it’s worth writing up or letting go. It’s a low-cost way to keep a pipeline moving without committing months to every stray thought.

  

Final thoughts

This post is two months late. I kept waiting to make it perfect, to have every project neatly scoped, every question precisely framed and keep thinking about what I’ll really be doing. Then I realized that’s not how any of this works. The whole point of pivoting into a fast-moving space is that the plan will change. Some of these projects will go nowhere. Others will turn into something I can’t anticipate yet. What matters is writing it down, putting it out there, and starting the conversation.

So if any of this resonates, whether you’re a potential collaborator, a student looking for a research direction, or just someone thinking about similar questions, I’d love to hear from you.

But here we are, 2 years later, and functional agents are here. Tools like Claude Code, OpenAI's Codex, and Gemini CLI have transformed what a single researcher can accomplish. This is the biggest step change in how AI can be used since reasoning models showed up. These systems read and modify files on your computer, orchestrate complex data pipelines, write and execute code, iterate on problems across dozens of steps without losing context. At a basic level, they save you from copy-pasting into a ChatGPT window. But that alone, combined with having your entire project in context, supercharges what's possible.

I’ve spent the past year using these tools for my daily research tasks. In that time, I’ve built scrapers, dashboards, and analysis pipelines that would have previously required months of work. I’ve run visualizations on datasets I collected (just say to qualitatively inspect them), iterated on research questions faster than I ever thought possible, and shipped single-authored papers I never imagined I’d have the bandwidth to complete. The productivity gains are amazing but they also unsettling in ways I’m still processing.

This isn’t another piece about how AI will transform research in the abstract (see my piece on that last year). This is about what it’s already doing to my workflow, to my students (or lack thereof), and to the uncomfortable questions I now face about what constitutes meaningful work in an era where execution has become cheap.

Let me try to synthesize what I’ve learned from this year of working with coding agents:

1. Too much academic discourse treats AI as a taboo to be avoided or a fad to be ignored. We need to take our heads out of the sand: these tools are becoming a competitive necessity, and pretending otherwise is a disservice to the next generation.

2. The parts of research that were already somewhat automated, like data analysis, visualization, literature review will (and have) become radically easier. The parts that were always hard and human like data collection, ethical judgment, asking the right questions still remain as hard as they were.

3. The economics now favor substituting AI for junior scholars on exactly the tasks that used to train those scholars. This is rational at the individual level and potentially catastrophic at the collective level. I don’t know how to solve this, but I think we need to talk about it more honestly.

4. AI helps you produce more, but the question is whether it helps you produce better. The 80% of mediocre work is expanding; the top 10% remains as hard to reach as ever. The researchers who thrive will be the ones who use AI to be more selective, not just more prolific.

5. If data analysis is free, the value shifts upstream to the questions you ask and the data you can uniquely access. For researchers who can develop their expertise in this space, this is actually an opportunity. But it requires recognizing that the game has changed.

6. I’ve benefited from these tools while contributing to dynamics I find troubling. I suspect most researchers using AI are in a similar position. The least we can do is be honest about the trade-offs.

Impact on PhD students

This coming year, I do not plan to admit any PhD students. There are many usual reasons like funding cuts, institutional pressures, and the usual constraints. But there’s another reason I’ve been reluctant to articulate: I felt like I didn’t need one. Not in the way I would have two years ago.

Most of the work a fresh PhD student does requires a lot of hand holding and training. This work now takes me a few minutes/hours with Claude Code/Gemini CLI/Codex. What would have been a semester-long project for a student learning the ropes can now be done in an afternoon. I needed a dashboard for a recent project, which I am sure would be a multi-semester Master’s project was built in a day. So many examples like this.

Joshua Gans and Andy Hall have written about AI-augmented research productivity, and a recent Science study shows LLM users publishing 30–90% more papers. But these “positive” takes largely benefit people like me—researchers who are already trained, who already possess the judgment to distinguish a good idea from a bad one.

I can frame this as efficiency, especially for someone early career like me... more time for thinking, for writing, for the “real” intellectual work. And there’s truth in that framing. But I keep feeling uncomfortable. Even though many might not express it, this is bound to happen and when it happens at scale, this will have a huge impact on our PhD training pipelines. Most of this ‘grunt work’ I am automating was a great way for students to experience the good and bad of research.

What’s a solution to this? Chris Blattman recently suggested that research assistants will shift to other tasks now that programming and annotation are increasingly solved. Maybe. I want to believe that. But I don’t think the transition will be clean. There will be turbulence, at least in the short term, as we figure out what those “other tasks” actually are and how people learn to do them.

Impacts on Research Quality

How do these tools then impact research quality? Luckily there’s evidence for this already. The productivity boost is real. The Science study by Kusumegi et al. showed that researchers using LLMs publish 30-90% more papers than their peers. But the same study found that traditional quality signals are decoupling from actual quality. Papers written with LLMs score higher on linguistic complexity, yet those high-scoring LLM papers are less likely to be accepted than high-scoring human-written ones. (note though the Science study is about writing and not just analysis, which is a different ball game altogether).

The tools make it trivially easy to produce work that clears the bar of “not obviously bad.” They do not make it easier to produce work that is obviously good.

Consider the distribution of academic work. In any academic conference, there’s a small set of papers (maybe 5-10%) that are obviously excellent. Everyone agrees they should be accepted. There’s another small set (another 5-10%) that are obvious rejects. The remaining 80% is where most papers are. These papers are usually competent but not exceptional, methodologically sound but not particularly novel, mostly incremental contributions that might or might not clear the bar depending on who reviews them and conference capacity.

AI is flooding the 80%.

Most of us, if we’re honest, live in the 80%. I certainly do. The question AI forces is whether we use it to produce more work in that middle band, or whether we use it to escape into the top 10%.

The temptation is to produce more. Analysis is no longer a moat. Typical academic data analysis tasks like running regressions, generating visualizations, and iterating on findings which once took weeks now take hours. The friction that used to slow us down is gone which increases volumes. A Nature study analyzing 41 million papers found exactly this pattern. AI-assisted research is clustering around popular topics where large datasets already exist. The authors find that early advances attract more researchers to the same problems, and the feedback loop concentrates attention rather than diversifying it. AI papers covered 4.6% less intellectual territory than conventional studies.

The AI tools optimize for throughput, and throughput is what we measure, so throughput is what we maximize. But throughput in the 80% is just noise because there’s so much randomness. It crowds the literature, burdens reviewers, and adds little to knowledge. The researchers who matter will be those who resist this pull.

AI-Proofing Research

Things will be different, and our training (both for students and ourselves) must reflect that difference. We cannot continue to treat data cleaning and basic coding as the primary learning experiences in a PhD. Instead, the value must shift upstream to the parts of research that have a “pro-human bias”. If execution is cheap, we must think about the parts of the funnel which are harder.

Think of questions upstream to the data analysis, like finding research questions are worth asking, interesting datasets to collect and how to collect them. Dean Ball argues that proprietary data access will become an even greater differentiator than it already is. I think he’s right, but the point generalizes: any input that is hard to acquire becomes more valuable when the processing of inputs becomes cheap. Things like novel data collection, deep domain expertise (or interdisciplinary collaborations), access to populations that don’t show up in existing datasets, etc.

AI removes friction from execution, which means the selection problem gets more important. The temptation to publish increases precisely when the value of publications decreases.

I really hope we use the time AI provides us to break into the top 10% making use of AI’s productivity gains as an opportunity to be more selective, not less. We need to use the time savings to think harder about what questions matter and recognize that the game has changed: analysis is cheap, so the advantage shifts to everything that comes before it.

What Academia Needs to Reckon With

I still see many academics resisting AI, dismissing it as a fad or talking about keeping research “pure”. I understand the instinct. But I think that resistance will increasingly become a luxury. People who don't adopt these tools will fall behind.

AI is now a productivity tool. That’s it. People who use it will be more productive than people who don’t. The Science study showed 30-90% increases in output. You can have opinions about what this means for scholarship, but you cannot pretend it isn’t happening. For PhD students entering the market (in certain fields) in the next few years, not using these tools is not a principled stance. It’s a competitive disadvantage.

For those of us who are tenured or approaching it, the stakes might be lower. We can afford to be thoughtful, to experiment, to wait and see. We built our careers under the old rules and we’ll most likely be fine under the new ones. But the next generation may not have that luxury.

Writing this piece has forced me to confront tensions I’d rather avoid. I benefit from these tools. I made a decision this year not to train a student who would have learned valuable skills through work I now outsource to AI. I’m contributing to dynamics I find troubling while also riding them to my own advantage. That’s uncomfortable to admit, but it’s true for most researchers using AI right now, whether they say so or not.

What I think academia needs, broadly, is to stop treating this as an individual ethics problem and start treating it as a collective one. The efficiency gains are immediate and visceral. The long-term costs are diffuse and speculative. It is rational to adopt these tools. That’s precisely why the problems I’ve described will keep getting worse: no individual has an incentive to slow down.

Final Thoughts

Even if AI gets really good, the bottleneck remains our time. I still have to look at the analysis and decide what to do with it. Even with completely functional AI scientists, someone has to look at the outcome and make the call. The learning for younger researchers might just lie in identifying which paths are worth pursuing. The PhD model might not disappear but we need to work on it to shift to the parts that actually require human judgment.

I don’t know what research looks like in five years. I don’t know whether the academic pipeline will adapt. I really hope it does. But I know that the agents have arrived, and this is indeed a step change in capabilities. The question now is what we do with them, and whether we can be wise enough to use tools this powerful without losing something essential in the process.

• Despite apocalyptic rhetoric about misinformation threatening democracy, there’s no grassroots demand for the solutions we’ve proposed, even from people who believe misinformation is a crisis. That’s a signal worth taking seriously.

• I think this is because Misinformation research has focused on process and not outcomes. We’ve built impressive infrastructure for fact-checking but we measure outputs (claims checked, labels applied) rather than outcomes (beliefs changed, harms reduced).

• This happened for a reason: outcome-focused research requires deciding what people should believe, which is a political judgment, not a scientific one. Process lets us avoid those hard questions.

• If we want people to care about misinformation the way they care about climate, we need to offer solutions that feel like serious responses to serious problems, not procedures that produce legitimacy without accountability.

Here’s something that’s been bothering me.

If you listen to how researchers, journalists, think tanks talk about misinformation, you’d think it was one of the defining crises of our time. In fact, the World Economic Forum thinks mis/disinformation is one of our biggest risks, two years in a row! Seriously… You regularly hear that misinformation is “killing people,” undermining democracy, tearing apart the social fabric. The rhetoric is apocalyptic.

If it is as dangerous as conflict and climate (at least according to the World Economic Forum), why does no ordinary person seem to care? Millions of people took to the streets for climate action. When people believe something is an existential threat, they mobilize. They demand solutions.

But nobody’s marching for asking for governments to prevent misinformation. There’s no grassroots movement demanding more content moderation. Even among people who genuinely believe misinformation is a crisis (our own side, so to speak) there’s a strange absence of enthusiasm for the solutions we’ve proposed.

I don’t think this is a marketing failure. I think it’s a signal we’ve been ignoring.

Lawyers all the way down

I’ve been reading Ezra Klein and Derek Thompson’s new book, and one observation stuck with me: of the last ten Vice Presidents (except Tim Walz) were lawyers. The authors points to something about how we’ve come to approach problems in American governance, and, they argue, in many other fields too.

Lawyers are trained to think in terms of process. Due process, proper procedure, compliance, defensibility. The question isn’t just “did we get a good outcome?” but “did we follow the right steps?” There are excellent reasons for this in law. Procedural protections matter. But according to Klein and Thompson, when this mindset colonizes everything from policymaking, regulation, institutional design, you end up optimizing for defensibility rather than results.

They argue this helps explain why so much of liberal governance feels like it’s going through the motions. Elaborate environmental review processes that take years, then block the solar farm anyway. Housing policies with all the right procedural safeguards that never actually produce housing. The process becomes the point.

Arvind Narayanan has made a similar observation about AI fairness research. We’ve developed sophisticated procedural criteria for determining whether an algorithm is “fair”: things like equal false positive rates across groups, calibration requirements, various mathematical definitions of non-discrimination. What we don’t do is ask whether deploying these “fair” algorithms actually makes anyone’s life better. Most AI “fairness” work is just trying to define fairness as a simplified, tractable optimization function and doing… optimizing for a function. A model can satisfy every fairness criterion and still leave the world worse off. But it’s procedurally clean.

I’ve started to think misinformation research (my own field) has gone through a similar trajectory. We’ve built elaborate processes for content moderation and fact checking. What we haven’t done, at least not convincingly, is show that the infrastructure produces the outcomes we care about.

The fact-checking industrial complex

Consider fact-checking.

Over the past decade, we’ve constructed an impressive apparatus. There’s the International Fact-Checking Network with its code of principles and certification process. There are partnerships between fact-checkers and major platforms. There are standardized rating systems, methodological guidelines, training programs. It looks serious and scientific and rigorous.

But what does it accomplish?

We can count outputs easily enough: claims checked, labels applied, articles published. What we can’t show, because the evidence is genuinely weak, is that any of this changes what people believe at scale. The research on fact-checking efficacy is decent, but it’s mostly lab-based studies and we really don’t know much about the long-term value of it. Some find backfire effects in certain populations. The overall picture is that fact-checking is a minor intervention in a complex system, and not a practical solution.

And yet the infrastructure keeps growing (until recently). Why? Because the process is fundable, visible, and defensible. Platforms can point to their fact-checking partnerships when held accountable. Researchers can publish papers measuring fact-checking’s properties. Foundations can report on how many claims were checked. Everyone has metrics to show. The metrics just don’t measure whether the thing is working.

For platforms, fact-checking has become a convenient way to deflect responsibility. Facebook funds over a dozen fact-checking organizations in India alone. Are they measuring whether these partnerships actually reduce belief in false claims? Whether the corrections reach the people who saw the original posts? Whether any of it changes downstream behavior? Not really. The funding is a rounding error in their budgets, a line item that lets them say they’re “doing something.” And for the fact-checkers, many of whom are IFCN-certified and dependent on platform money, it is a business model. I don’t blame them for taking the work. But we should be honest that the incentives here are about visibility and legitimacy, not demonstrated impact.

Content moderation has followed a similar trajectory. After years of criticism, platforms developed elaborate procedural responses: appeals systems, oversight boards, transparency reports, detailed policies with carefully defined categories. These are all procedural solutions. They make the process more legible and defensible. What they don’t do, and more importantly, what nobody really asks, is whether the underlying harms decreased. The Facebook Oversight Board is a perfect example. The board handles maybe a few dozen cases a year out of billions of content decisions. Its impact on what actually happens on the platform is negligible. It’s only a procedural entity. Everyone knows it, no one says it.

Why we ended up here

There’s something deeper going on here, though, and I don’t think it’s cynicism or grift. I think we retreated to process because the alternative is genuinely hard, maybe impossible.

Philosopher Dan Williams has written a provocative essay arguing there can’t really be a “science” of misinformation. His argument goes like this:

If you define misinformation narrowly (something like demonstrable falsehoods, clear-cut lies) then it turns out to be relatively rare. Most people don’t encounter much of it. And the people who do consume a lot of it are a specific minority with pre-existing characteristics: conspiratorial worldviews, institutional distrust, strong partisan animosity. They seek out misinformation because they already distrust mainstream sources, not the other way around.

So we slightly broadened the definition to include not just false claims, but also “misleading” content more generally. True statements taken out of context, selective presentation of facts, technically accurate but deceptively framed information.

The problem is that this category is essentially boundless. All communication is selective. Every news article chooses what to include and exclude. Every framing emphasizes some aspects over others. Is wall-to-wall coverage of rare vaccine side effects “misinformation”? What about wall-to-wall coverage of rare police shootings? Both involve accurate information presented in ways that might skew people’s sense of base rates. Who decides which selective emphasis is misleading and which is just journalism?

Williams’ uncomfortable conclusion is that answering this question requires deciding what people should believe, which isn’t a scientific judgment but a political one. There’s no neutral vantage point from which to determine that some true-but-selective claims are “misleading” while others are legitimate reporting. Your answer will inevitably reflect your prior beliefs and values.

This, I think, explains the procedural turn. Outcome-focused misinformation research would require us to specify what outcomes we’re trying to achieve. Less belief in false claims? Which false claims, and what if it’s not possible to decide whether they’re false? Better “epistemic health”? What does that mean, and according to whom? Less polarization? That requires a theory of what the right amount of polarization is.

These are hard questions, and reasonable people disagree about them. Process lets us avoid them. We can study the properties of content, measure sharing patterns, evaluate the internal consistency of fact-checking methodologies, all without ever specifying what success would look like. It feels scientific precisely because it sidesteps the value-laden parts.

What we lose

The cost of this procedural focus is significant, in my opinion.

First, like we did with fairness, it is easy to think of solutions as an abstraction and we end up treating symptoms rather than causes. If misinformation consumption is concentrated among people who already distrust institutions, then fact-checking the specific claims doesn’t address the underlying distrust. The distrust has sources (some of them legitimate), and those sources aren’t going away because we label a few Facebook posts as “missing context.”

Second, we build interventions that nobody wants. The lack of grassroots demand for our solutions is a problem and we should acknowledge that. I think people intuitively sense that fact-checking is a technocratic non-solution. It addresses the surface manifestation of a problem while leaving the deeper dynamics untouched. When someone doesn’t trust the New York Times, having a fact-checker affiliated with the New York Times tell them a claim is false isn’t going to land. We know this. We keep doing it anyway.

Third, we’ve created a system that produces legitimacy without accountability. Platforms can point to their trust and safety teams, their fact-checking partnerships, their transparency reports. Researchers get grants and publish papers. Think tanks produce frameworks and case studies of how bad things were for a specific event. Everyone has something to show for their work. But nobody has to demonstrate that beliefs changed, behaviors shifted, or harms decreased. The process justifies itself.

So what now?

What would it look like to take outcomes seriously?

I don’t have an answer, but there’s one open (and really important) problem that has existed for over a decade for which we dont have a clear answer.

We should focus on behaviors rather than beliefs. It’s genuinely hard to know what people “should” believe, and beliefs are difficult to measure anyway. But behaviors are more tractable. Did exposure to certain content cause people to act differently, e.g. to not vote, to take a medication against medical advice, to show up at the Capitol? Of course this is not easy to operationalize as easily as many other problems like developing accuracy nudges or inoculation or fact checking.

We might be more honest about limits. Maybe there’s no technocratic solution to misinformation. Maybe it’s fundamentally a political problem that requires political responses such as changes to media ecosystems, economic structures, democratic institutions. If that’s true, the honest response is to say so, not to keep building elaborate procedural infrastructure that makes us feel like we’re doing something.

Obligatory caveats

I want to be clear about what I’m not saying.

I’m not saying misinformation researchers are cynical or that fact-checkers are useless or that platforms shouldn’t moderate content. I’m saying that we’ve built a field optimized for defensibility and process, and we’ve avoided hard questions about whether any of it matters.

This isn’t a comfortable position for me. I’ve spent years working in this space. I’ve written papers using the standard frameworks, gotten grants from the standard funders, participated in the standard conversations. I’m complicit in what I’m criticizing.

But if we actually believe our own rhetoric, that this stuff is important, that democracy is at stake, then we owe ourselves some honesty. The elaborate infrastructure we’ve built may not be working. The reason nobody’s marching for fact-checking might be that people can sense, even if they can’t articulate it, that we’re going through the motions.

We need to realize that we’ve focused on proceduralism. We should also answer seriously the question that process lets us avoid: what would it actually mean to succeed? And being honest if we don’t know, or if the answer is that success isn’t really up to us.

I started by asking why nobody in the public is asking for a solution to misinformation. I think the answer is that people care for things they believe will actually impact their lives. Climate protesters believe that policy changes can avert catastrophe. The connection between action and outcome felt real. We haven’t given people that. We’ve given them labels on Facebook posts and oversight boards that review a dozen cases a year. We’ve given them processes.

But if we ever want people to care about this the way they care about climate, we have to offer them something worth caring about. Not procedures. Outcomes.

A few more caveats, because this is the internet:

There are many obvious oversimplifications through out the post. When I say “we haven’t demonstrated impact,” I don’t mean literally zero people have tried. Some have.

I am an active member of this community and still believe misinformation is important. That doesn’t mean I don’t get to have opinions about how we study it.

It’s obviously easier to poke holes at others’ work than to do the work. I really don’t mean to do that here. I’m trying to articulate something I think many of us (i.e. at least 2 people) feel but don’t say out loud.

Yes, there are bad fact checking companies, running fact-checking outfits primarily as businesses, researchers chasing fundable topics rather than important ones. But not everyone is a bad apple. The structural incentives are the issue, not individual moral failings.

This is not an academic article. I know for sure others have made similar claims, but I’m not doing an extensive literature review here. Omission is not out of malice. I do not claim that all these ideas are mine.

Why Video Needs AI

Video has become the dominant medium of our daily information diet, yet it remains remarkably difficult to analyze at scale. The complexity stems from video’s inherently multimodal nature: it’s an interplay of text (what was said), audio (how it was said: tone, volume, emotion), and visual frames (who said it, their expressions, their screen time). Understanding video requires untangling all three simultaneously.

The scale of video content is also huge and needs automated tools. A single year of prime-time opinion programming on just three U.S. cable networks generates over a billion words of caption text. For decades, we’ve had an enormous archive of television broadcasts sitting largely unanalyzed. The Internet Archive alone contains thousands of hours of cable news programming with closed captions. Stanford and Vanderbilt’s TV data collection spans years of broadcasts. This data has always been available, but never truly analyzable. What we’ve lacked until now is the ability to segment broadcasts into meaningful units (who said what, when), identify speakers and their political affiliations, and classify the stance of each exchange.

AI has finally made this tractable. Below are two applications from my lab that showcase what’s now possible.

Application 1: Measuring the Disappearance of Debate on U.S. Cable News

This post is about our recently published paper titled “Disagreement is Disappearing on U.S. Cable Debate Shows”, to appear at ICWSM 2026. The paper can be found here.

American prime-time cable news is a big deal. It’s where millions of Americans get their political information. Shows like Tucker Carlson Tonight, Hannity, and The Rachel Maddow Show routinely draw 3-4 million viewers per night, and the top-rated Fox opinion shows rank among the most-watched programs on all of cable television. These broadcasts occupy the most valuable real estate in the TV schedule: the 8-11pm block when viewership peaks and advertising rates spike.

The influence of these shows extends far beyond ratings. We know for sure Trump watches them but there are countless instances of lawmakers scheduling hearings around them, and social media influencers amplifying these clips. In general, the people who watch these shows are also more likely to vote, so their influence is much broader than the 3-4 million viewership (which is much smaller compared to say a massive social media influencer).

Many of these prime-time shows on Fox, MSNBC, and CNN brand themselves as “debate” shows. The format promises competing viewpoints: a host who interviews or spars with guests about the political story of the day. A true debate means you have competing views on screen—each side presents a different perspective, they argue about it, and the audience decides. We wanted to answer the question: How much real “debate” is actually happening in these shows?

To measure this, we look at two things. First, who gets invited: are shows booking guests with genuinely different viewpoints, or just friendly faces? Second, what happens on air: when host and guest interact, do they actually disagree? We use host-guest disagreement as a proxy for real debate. If a show consistently features like-minded guests who nod along with the host, that’s not a real debate.

We assembled a corpus of 21,000 episodes (roughly 8,000 hours of video) from 24 flagship opinion shows on Fox News, CNN, and MSNBC, spanning 2010 to 2024.

The technical pipeline involved several AI components. First, automatic speech recognition and speaker diarization using Whisper to segment each broadcast into host and guest turns. This identifies who said what and when. Second, a high-fidelity stance classifier built through an iterative process combining GPT-4 seed expansion with crowdsourced annotations and fine-tuned open-source models, achieving 89% accuracy on detecting agreement, disagreement, and neutral exchanges between the host and guest. Third, guest identification and political affiliation coding using LLMs to extract names from opening monologues, cross-referenced against the a campaign finance database for. Finally, we used BERTopic for identifying the topics discussed in each episode.

We obtained 2.13 million speaker turns, each labeled with who was speaking, what they were discussing, and whether genuine disagreement occurred.

I want to emphasize how reliable this pipeline is. We can identify not just the text of what was said, but the speaker, the topic, the stance, and do this across 15 years of programming. This level of structured analysis of video content simply wasn’t possible before.

What we found

Disagreement is rare, and getting rarer. Across 2017-2024, barely 15% of host-guest exchanges register as disagreement. MSNBC is lowest at 13%, Fox highest at 17%. More troubling: on Fox and CNN, the share of disagreement turns falls sharply each year. Tucker Carlson Tonight, for instance, started with roughly one-third of exchanges coded as disagreement in 2017 by the time the show ended in 2023, it had dropped to about 15%.

The decline is historically contingent. This wasn’t always the case. Data from the early 2010s show Sean Hannity operating at nearly double today’s disagreement rate. The “debate” format once accommodated much more disagreement and debate.

Partisan gatekeeping amplifies the trend. The channels not only book ideologically friendly guests but also moderate them differently. Republican guests face the stiffest pushback on MSNBC; Democrats get the gentlest treatment there. The mirror pattern holds on Fox. CNN has narrowed toward the middle.

The most polarizing topics attract the least disagreement. The culture war flashpoints like abortion, gun rights, immigration, that dominate coverage, show the lowest levels of on-air disagreement. These segments, most likely to mobilize audiences online, have become echo chambers on air.

What this means

The programs that bill themselves as “debate shows” have largely stopped debating. By 2024, the combined share of agreement and neutrality exceeds 90% of airtime across all networks. Meaningful counter-argument has become the exception.

This matters because television still has a ton of influencing power, and often elite rhetoric and social media conversations frequently echo televised talking points. If the “debate” has quietly become one-sided affirmation, the viewers see their side’s view as the “settled truth” because they rarely are challenged on air.

There is a really nice connection here to the political science theory of cross-cutting cleavages. The theory basically says that political conflict gets dangerous when your party identity becomes the singular line that divides you from your neighbor. With the disappearence of disagreement, these shows hide all the other ways we might connect or differ. Instead, they present a flattened, binary world where everything is just Red vs. Blue.

Application 2: Multimodal Analysis of Incivility in Indian TV Debates

This work appears in our paper "Television Discourse Decoded: Comprehensive Multimodal Analytics at Scale", published at KDD 2024.

The U.S. cable news study focuses primarily on what was said: transcripts, speaker turns, stance detection. But video offers so much more. In this work, we developed a comprehensive multimodal toolkit that goes further, analyzing Indian prime-time television debates across text, audio, and video simultaneously.

India's prime-time TV debates are infamous for their chaos: shouting matches, overlapping speech, dramatic confrontations. They're watched by millions and have been criticized for compromised journalistic integrity and excessive dramatization. But no one had quantified any of this at scale.

We processed over 3,000 videos (totaling 2,087 hours of footage) from one of India's most-watched English news debate shows (We're not naming the show; given the current media environment in India, an abundance of caution seems wise for my collaborators based there.). The pipeline goes beyond simple transcription to extract granular signals across all three modalities: from audio, we detect overlapping speech (when multiple people talk over each other) and shouting; from video frames, we identify faces, track screen time by gender, and measure how much visual real estate each speaker is allocated; from transcripts, we extract panelist names, match them to political affiliations, and analyze how language is deployed differently when discussing different political actors.

What we found

The data confirms what critics have long suspected, but now with hard numbers. Incivility is pervasive: on average, 9% of debate time involves shouting, and in contentious topics like Kashmir, Religion, and the Citizenship Amendment Act, over 20% of airtime features overlapping speech, people literally yelling over each other. For comparison, we ran the same pipeline on debate shows from the U.S. (Morning Joe, Presidential Debates), UK (Sky News), and France (France 24). Indian debates showed statistically significantly higher incivility across the board.

Gender representation is also stark: Women account for just 7.5% of total screen time across all videos, and in debates about national security topics like the Pulwama terror attack, that drops to 5%. Even when women do appear, they’re given less visual prominence: male faces occupy an average of 1.5 times more screen space compared to women.

There is also a systematic bias in guest selection and framing. Pro-ruling-party panelists are invited at roughly 3-to-1 ratios across most topic categories. And the language tells its own story: when we analyzed which words predict whether a sentence is about the ruling party versus the opposition, the ruling party is associated with “victory,” “development,” and “Modi wave,” while the opposition gets “dynasty,” “shame,” “fake,” and “lie.”

What this means

The benefit of this kind of quantitative analysis is that we can take a qualitative hunch that many have long held (that these shows are biased and sensationalized) and demonstrate it systematically across thousands of hours of programming.

The numbers paint a picture of a media environment that has abandoned the pretense of balanced discourse. When 9% of airtime is shouting, when women get 7.5% of screen time, when ruling-party voices outnumber opposition 3-to-1, and when the very language used to describe each side is systematically different, it’s really a failure of “journalism”.

This matters because India is the world’s largest democracy, and television remains enormously influential: the show we studied draws over five million daily viewers. When that many people are exposed nightly to a media environment where one side is “victory” and the other is “shame,” where debate means shouting and women are nearly invisible, it shapes what citizens come to expect from public discourse itself. The normalization of incivility on screen licenses incivility off it (this is harder to prove causally, but seems likely).

Looking Ahead

We started this line of work with TV data because it’s the “cleanest” form of video with professional audio quality, stable visuals, structured formats. It was a good testbed for pushing multimodal pipelines to their limits. But the tools generalize (with caveats). We’ve also been experimenting with vision-language models on messier, social media videos and images from WhatsApp. The results are promising, at least for certain tasks. I’ll write about that work in a future post.

I'm proud of both the papers discussed above, but what excites me most is that these phenomena are now measurable at all. Once you can measure something, you can track it, over time, e.g. Is disagreement on cable news declining year over year? You can compare across contexts: Why are Indian debates so much more uncivil than French or American ones? You can ask questions that weren't even possible before. That's what opens up when video becomes data.

This is the through-line connecting all the work I’ve been writing about. print newspapers, TV broadcasts, video heavy social media. These datasets have existed for years, sometimes decades. What’s changed is that AI has crossed a threshold: it’s now reliable enough, at scale, to transform media that was always available into data that is actually analyzable.

Next up: I’ll write about our experimental work using AI to process Google Street View images, trying to connect what we observe about people’s online behavior with the physical conditions of where they live.

The relationship between the news coverage and advertisers has long been a subject of interest in media studies and communication. For many decades, the separation between editorial content and advertising revenue was considered the a key pillar of journalistic integrity. However, as the economic foundations of traditional media (particularly in India) have shifted, the question of whether this firewall still stands has become more urgent. Our recent research, answers this question: Does Ad Money Influence the News?. More important (IMO) than answering the question, the paper provides a key application of AI tools to enable the large scale study of print newspaper content (PDFs), something which was not possible previously. This enables us to extract and estimate interesting things like ad spend amounts by corporates and governments on top newspapers in the country, e.g. an interesting finding is that we estimate government and corporate spending on ads to be around 1 Billion USD each (around 9000 crores).

Who Cares about Print Newspapers?

The first question I had about this problem was print newspapers are a thing of the past and most news consumption is digital. I came across some interesting stats. Even in 2025, print news is highly trusted, and has a huge marketshare in terms of advertising revenue. This is true all over in both US and India.

While it is true that print news is dying and digital landscape is often considered the primary battleground for media influence, print newspapers remain a formidable force, particularly in markets like India. Globally, print continues to command a higher level of trust than news websites, and in the Indian context, it represents the world’s second-largest market with over 110 million copies sold daily. Despite this influence, print media has historically been a “black box” for researchers. Unlike digital media, where data can be scraped and analyzed with relative ease, print archives are often proprietary, expensive to access, and rarely digitized at scale, especially for non-Western languages. The figure below shows how advertising on print has pretty much remained constant over the last 8 years (except covid).

AI to the rescue

To understand trends in print advertising, we developed a novel AI-driven pipeline designed to ingest and interpret the physical layout of a newspaper. Utilizing YOLOv8s for image segmentation, we achieved a 96.8% precision rate in identifying and separating articles from advertisements. By combining this with advanced Optical Character Recognition and the IndicTrans2 translation model, we were able to process over 12,000 newspaper editions spanning six years and three languages. This technical siege allowed us to transform hundreds of thousands of individual pages into a structured dataset ready for rigorous economic analysis.

I cant stress how good and reliable this is. We can not only identity the article text (in arbitrary column width and shapes), but also if something is an ad or a news article, who the advertiser is and the area taken by the ad. This level of AI which is practical, dependable and useful will really open up a ton of possibilities for analyzing all the historic PDF documents, hopefully breaking the “PDF is where documents go to die” meme. We already see amazing use cases of this, e.g. Melissa Dell’s research using similar AI tools (our work was inspired by their tools) to process historic newspapers from the Library of Congress.

Government vs. Corporate advertising

We quickly found that there are two big spenders in this space who make up over 60% of the ad spends - big corporates and Government. Next we mapped out the strategies employed by these two players. Corporate advertisers operate with a clear emphasis on high-impact visibility. Their playbook is characterized by a pursuit of premium real estate, with nearly 28% of their total ad area concentrated on the first, third, or last pages of the paper. They favor large, standardized formats like quarter, half, and full-page spreads, mostly designed to grab the reader’s eye and signal brand authority.

In contrast, government advertising follows a logic of mass dissemination rather than high-impact branding. Government ads are often small, text-heavy, and dispersed throughout the “inner” pages of the publication. Over 88% of government ad area is found away from premium spots, reflecting a mandate focused on public notices, tenders, and legal compliance. While corporations buy billboards within the paper, the government posts bulletins. This distinction is crucial because it sets the stage for how these two entities interact with the news cycle itself.

Impact of Advertising on News Sentiment and Volume

To answer our key question of whether ad spend gives you positive (and more) coverage, we set up a regression to determine if ad spending correlates with how a company is covered in the news. The results for the corporate sector were both clear and statistically significant. We found a robust positive correlation: as a company increases its advertising expenditure, the volume of news coverage it receives rises, and more importantly, the tone of that coverage becomes markedly more favorable.

Quantitatively, a 1% increase in a company’s weighted ad ratio is linked to a 0.0189-unit increase in the total sentiment score of its news coverage. On a sentiment scale ranging from -1 to +1, this represents a substantial shift. This finding suggests that for corporate entities, advertising serves as a “reputation insurance” to ensure a friendly editorial environment. The economic reality is that for-profit newspapers, which rely on advertising for 60-70% of their revenue, face an inherent pressure to maintain positive relationships with their most significant financial backers.

For governments, its not as clear. We went in with the hypothesis that if corporate spending buys favor, government spending (which often exceeds corporate expenditure) would do the same. However, our analysis revealed a surprising relationship for government advertising. Increased government ad spending is actually correlated with more negative coverage of government entities.

This paradox can be explained by the lack of “transactional incentive.” Because governments are often legally mandated to publish their notices in all major newspapers, the revenue they provide is essentially guaranteed by law. Unlike a corporation that can pull its ads if it dislikes a headline, the government’s spending is largely non-discretionary. This removes the newspaper’s incentive to “soften” its coverage.

Implications

To be honest, I am less excited about the findings of this specific paper and more excited about the tools and what can be done with them. I know this is a trap many of us fall into, where we go looking for nails because we have a hammer, but I think this time it is different. I really think the next few years is going to be a golden age for CSS research (as I wrote earlier). The fact that these AI tools are now SO good to be useful and applicable really unlocks a ton of possibilities for such public interest research. We are now extending this work to study coverage bias in print and online newspapers. For example, we want to understand what choices editors make when they have to fit content into the limited space of a print newspaper, a constraint that does not apply in digital news.

By open-sourcing our code and dataset, we hope to provide a new toolkit for media transparency. Hopefully, there will be more work in this space, more broadly on parsing and extracting structured text from PDFs. I will over the next few weeks talk about our other work processing videos, images and audio with AI tools, which can also lead to great insights from datasets (e.g. television data) which were always available but hard to process automatically.

Note: the figures used in this post were generated with NotebookLM. Their “Slide Deck” feature is REALLY good.