You tried ChatGPT, asked it write an email and then stared at the output, thinking, “This sounds like a LinkedIn post written by a bot who just discovered the word leverage.”

Or you’ve been around enough to hear everyone talking about AI agents, vibe coding and the future of work, but then you got overwhelmed and turned on Love is Blind. (We’ve all been there.)

Subscribe now

Why you don't need a $500 prompting course

It’s not that you don’t know enough about AI. AI, for the most part, is a text box that you fill in and a chat you interact with. There are a lot of people invested in making you feel like you need to pay $500 for a prompting course. My solve for that is: ask the AI to write your prompt. Done!

AI is changing quickly, but that doesn’t mean you need to chase every new model, tool or acronym. That’s the worst way to learn.

What you need is a good mental model. You need to experiment, and you need enough experience and judgement to know when AI is being useful and when it’s a confident liar.

There are 5 things you need to know. Today, I’ll walk you through all of them, and I’ll tell you what you can ignore (which is most of it).

I’m not an engineer. I studied journalism, worked in media, and now I lead AI communications at a big finance company. I learned everything in this guide the same way you’re about to.

The 5 steps:

1. Understand. Learn enough vocabulary to understand what people say in meetings.

2. Experiment. Pick a problem you’re wrestling with right now.

3. Question. Understand that AI can always be wrong, even when it seems right on the surface.

4. Verify. Check all the facts in your output.

5. Integrate. After you do all of that, it’s time to make AI part of your workflow.

The 5-step framework for AI literacy

Step 1: Learn enough AI vocabulary for the boardroom

Let’s start with a question that sounds basic, but stumps many of AI’s top minds.

Quick reference: Core AI terminology

At the broadest level, the term artificial intelligence refers to computer systems designed to perform tasks that we associate with human intelligence: recognizing patterns, predictions, understanding language, and solving problems.

Generative AI is the kind of AI you’ve probably been interacting with most. Gen AI systems can generate text, images, audio, video, and code.

A large language model, or LLM, is a type of AI system that generates language by learning patterns from enormous amounts of data.

Then there are reasoning models, multimodal models, AI agents, and AI assistants. And a hundred new terms coming at us every week.

You don’t need to memorize them all. But you do need to understand the basics.

The fundamental mental model: AI isn’t a search engine

Something AI is not: a search engine.

When you ask an LLM a question, it’s not necessarily retrieving a single correct answer from a database. It generates a response based on patterns.

That’s one reason AI can produce something that sounds completely convincing and is completely wrong. It’s also why asking the exact same question twice doesn’t necessarily produce exactly the same answer.

→ This Is Why AI Never Answers the Same Way Twice

Once you understand this, you’ll start interacting with AI differently. You’ll be less impressed by confident-sounding answers. You’ll get better at asking follow-up questions. And you’ll understand why verification matters.

Honestly, this is one of the most important pieces of AI literacy you can develop.

Step 2: Experiment on real problems, not hypothetical scenarios

You don’t need 100 secret ChatGPT prompts. You don’t even need a “master prompt.” And you definitely don’t need to spend your Sunday building a 47-page prompt library.

10 practical AI use-cases for daily work

Instead, start with something you already do. Ask AI to:

• Explain something you’re trying to understand

• Brainstorm ideas

• Summarize a long document

• Help you prepare for a meeting

• Give you a counterargument

• Turn your notes into questions

• Quiz you on something you’re learning

• Help you compare two approaches

• Identify gaps in your thinking

• Give you a first draft you can improve

The important part is what happens after you get the answer.

Read it. Question it. Edit it. Check it. Decide whether it’s any good.

That last part is where a lot of the magic, and a lot of the danger, lives.

Why AI is the ultimate learning accelerator

We’re used to talking about AI as a productivity tool. Write this. Summarize that. Make me a spreadsheet. Draft this email.

But one of the most interesting things AI can do is help you learn.

You can ask it to explain a difficult concept in plain English. Then ask for an analogy. Then ask it to quiz you. Then ask it to argue the opposite position. You can turn one conversation into a personalized study session.

That doesn’t mean everything AI tells you is correct, but it can be a surprisingly useful way to explore a subject.

If you’re a student, knowledge worker, lifelong learner or someone who’s ever opened a 40-page report and thought “absolutely not,” this one’s for you.

A tool worth trying: Gemini Notebook (formerly NotebookLM)

If you’re looking for a practical AI tool to experiment with, Gemini Notebook is a great place to start.

The idea is that you give it sources: documents, PDFs, notes, research. Then you use AI to explore that material. That can be surprisingly useful when you’re dealing with information overload.

I’ve used the same idea for everything from research to learning, and I wrote a guide covering 12 different ways to use it.

Don’t try all 12. Pick one that solves a problem you want to work on. That’s a much better way to learn AI than downloading five apps you’ll never open again.

Step 3: Train your instinct to detect AI mistakes

Prompting isn’t the most valuable skill when you’re using AI.

Obviously, how you communicate with an AI system affects the output. But the more time I spend using AI tools, I can report that memorizing elaborate prompts isn’t the move.

The valuable skills are things like:

Curiosity. Can you ask better questions?

Skepticism. Can you notice when something doesn’t smell right?

Judgement. Can you tell whether the output is actually useful?

Context. Can you explain what you’re trying to accomplish?

Step 4: Establish workplace safety guardrails

There are 3 things to get right early. As the great RuPaul often says, it’s easier to stay ready than to get ready. None of the following guardrails are difficult to set up and they’re all harder to fix later.

1. Fact-check high-stakes information

AI can be useful, but It can be extremely confident while being incorrect.

This is sometimes called a hallucination, although I think “the computer made something up and said it like it was absolutely certain” is a more useful description.

Get into the habit of verifying AI’s output, especially:

• Statistics

• Sources and citations

• Current events

• Financial information

• Legal information

• Medical information

• Anything else that affects an important decision

The more important the answer, the less you should rely on AI alone.

2. Protect your data privacy

This feels obvious, but I see people do it all the time.

AI tools are incredibly convenient, which makes it very easy to paste in whatever you’re working on without stopping to think about what you’re sharing.

Don’t give AI:

• Passwords and credentials

• Confidential work documents

• Personal financial information

• Private information about other people

• Proprietary company information

• Sensitive conversations

• Anything you wouldn’t want exposed

Privacy standards vary between AI products, account types and settings, so understand the tool you’re using before treating it like a private diary. But honestly, it was only a few weeks ago that Claude users discovered some of the chats they thought they were sharing with a friend or family member ended up on the public internet. The same thing happened last year with Meta’s AI chats. It’s better to be safe than sorry with sensitive private information.

3. Be careful with autonomous AI agents

A chatbot responds to your instructions. An AI agent can take actions on your behalf. That could mean navigating websites, using software, retrieving information or completing a series of tasks.

That sounds convenient until you remember that computers are very good at doing exactly what you asked them to do, including when what you asked them to do was based on a misunderstanding.

The more autonomy an AI has, the more important permissions, privacy, oversight and error recovery become.

Before you start giving an AI access to your inbox, files, accounts or any part of your digital life, read this:

My general principle: start with tasks where mistakes are reversible.

Share

Step 5: Integrate AI into your daily workflow

So you’ve learned the terms, done the experiments, learned to check when AI is lying, and put some guardrails in place. This is where you can start to integrate AI into your day-to-day.

I use AI throughout the process of making this newsletter. It helps me research, brainstorm, find gaps in an argument, pressure-test ideas, and work through drafts.

But AI is part of my process, not the author of my thinking.

I still decide what I believe. I still check important claims. I still make editorial decisions. And I still have to decide whether an idea is worth putting in front of you.

That’s the difference between “I’ve played around with AI” and “I know how to use AI.”

The big dilemma: Will AI dull our critical thinking?

If AI writes our emails, summarizes our reading, generates our ideas, researches our problems, and makes our decisions easier, what happens to our underlying skills?

I’m not suggesting we all throw our laptops into Lake Ontario and return to handwritten correspondence. But we should be intentional about what we delegate. I’m the first to say that I’ve delegated the wrong thing. And people notice.

If writing is part of your work:

AI can make writing faster. It can also make it sound interchangeable. Your judgement, taste, perspective, and voice matter now more than ever.

You’re never going to be caught up

This might be the most important thing I can tell you.

There will always be another model, app, or acronym. Another person on LinkedIn announcing that everything has changed, again.

Ignore most of that.

Instead, learn enough to understand what’s happening. Use AI enough to develop your own instincts. Question it enough that you don’t blindly trust it. And keep doing the thinking yourself.

Because the goal isn’t to become the person who knows the most about AI. It’s to become a person who knows how to live and work well in a world where AI exists.

That’s what Human+AI is about.

Welcome.

Subscribe now

P.S. If you want to go down the rabbit hole. (Of course you do.)

🛠 Want to build things? → The beginner’s guide to vibe coding

💻 Want to run AI locally? → Local AI: What it is, why it’s trending, and how to install it

I was standing at my kitchen counter eating Skinny Pop (my current snack of choice) that I didn’t remember starting to eat. I thought about the fact that I haven’t been to the gym in a long time. I’ve been slacking on meditation, too. And grounding, and my red light mask.

In fact, I’ve been slacking on taking care of myself at all lately, which lots of us can relate to from time to time. Life gets busy, you start taking shortcuts, you give yourself an extra hour of sleep instead of going to the gym.

Subscribe now

I know what I need to do. Strength training, eat more protein, sleep more, meditation, et cetera. A lot of apps shame you for not doing the thing. (I’m looking at you, you abusive narcissist, Duolingo!) I don’t need an app to tell me I’m not good at going to the gym, I can feel that in my bones. (Literally.)

So I got to work building a live, local app on my phone. One of the things I didn’t want to do was give a company all my health information. It’s built for my specific health concerns, tests against real usability findings, and I did it with Fable in Claude Code. I used my experience working in product design to run extensive research, write product requirements, design, protype, run (AI) user testing and a production deploy.

I’m going to tell you what I did, how I worked with AI with real constraints and standards, and how my part in it made it valuable.

Directing AI vs. vibe coding: What’s the difference?

Instead of prompt engineering, or following a template, I took what I knew about product development, and I product managed a team. My AI agent collaborators (researchers, designers, QA testers) contributed to the project.

“Prompting” implies you type a request and hope. Directing means you set the brief, hold the standard, review the work like an editor, and send it back when it’s wrong. Those are different roles, and only one of them produces something you’ll use.

My 5-step AI product management workflow

There’s a term for building software by describing what you want and letting AI fill in the rest: vibe coding. It’s fast and it’s real, but you could end up with something that technically runs but doesn’t get to the heart of what you’re trying to make.

Instead, I ran a product design life cycle:

• Discovery. Before writing the product, I had Claude Code interview composite users grounded in real research. A busy exec managing food allergies, a lifelong athlete hitting a wall, a nervous beginner scared of the gym, a stress-eater with zero time, and a woman with ADHD who struggles with executive function.

• A real product requirements document. Prioritized features, success metrics, explicit scope cuts.

• Design and a clickable prototype. I could tap through before the product was coded and built.

• Usability testing against my personas.

• Build, verify, and deploy. I created a real, installable web app, live on my own Cloudflare account, working on my phone right now.

I didn’t skip the work when I used AI. But I was able to go through a product build without an entire product team and massive budget. The tools allowed me to do the job of 10 people.

Share

Where AI excelled vs. where human judgement was essential

Where AI was breakthrough tech:

• Parallel research threads: Analyzed medical literature with citations, performed competitive teardowns of top health apps, and parsed the neuroscience of nervous-system regulation simultaneously.

• Automated QA and self-testing: The app opened a browser, tested its own user flows, and corrected calculation errors before human review.

• Low-cost experimentation: Rapidly prototyped features like cycle-phase tracking, nutrition recognizers, and science-backed calming techniques without sinking time or money.

Where the human was the whole game

• Refusing the obvious default. I can’t use an app that will shame me for eating a donut. AI would have built that for me based on a lot of what’s already out there. My app tracks the same numbers, but not as something I will fail at. AI’s nature is to build the average. Your point of view is what makes it novel and interesting.

• Catching what the research alone couldn’t. If you’re making something for yourself, you’re the end user, which means catering to yourself. I valued the usability testing because it surfaced things that I wouldn’t have thought of. But, at the end of the day, I made each call to create a custom app that worked best for me.

• Holding an editorial line. When I researched nervous-system regulation, I made sure everything I included for myself was backed by published research. I didn’t include trendy things that haven’t been proven, because I wanted my app to be effective.

• Knowing the user (aka me) in my bones. I’m the primary user. I have ADHD, food allergies, and I know what I’ll respond to (not Duolingo). Research works for the masses, but I know what works for me. That allowed me to iterate into something genuinely useful.

Limitations of AI app development

When the cost of making something gets cheaper, you can see how your most valuable resource becomes the judgement and taste that makes the output or product genuinely useful.

The skills that carried this build were decomposition, standard-setting, critical review, editorial judgement, and the ability to say no to a feature that was fine but wrong. Those are management and taste skills.

AI made mistakes I had to catch, like a broken piece of logic in how the app decided what to suggest.

It’s a fast, literal, occasionally-wrong collaborator, not an oracle. And there are real ceilings: the web version of my app can’t talk to my Apple Watch or send me a notification, because iOS blocks both for anything that isn’t a native app. Some things still need an engineer, or at least a different kind of build than I did.

Key takeaway: Management and taste are the new code

The barrier to building the tool you keep saying “someone should build” has mostly disappeared. What hasn’t disappeared — what’s actually gotten more valuable — is your ability to know what’s worth building, and the standards you’re willing to hold it to.

The literacy that matters now isn’t learning to code. It’s learning to direct: brief clearly, break the work into pieces, review it like an editor would, and refuse the version that’s technically fine but wrong. That’s a learnable skill, and it looks a lot more like good management than like engineering.

I built something that helps me, with a real need, in a way that would have been impossible two years ago. But it needed me to do it.

Want the app? Become a Founding Member and get it, along with help to personalize it for your needs.

Subscribe now

AI in the news

A mostly plain language primer on Anthropic’s new watermark and how it can be detected (Wondering about AI) Karen Spinner breaks down the new Anthropic statistical watermark into Claude-generated text. The watermark complies with the EU AI Act, joining Google and potentially OpenAI, with existing Claude models expected to be retrofitted by December 2026. Spinner’s experiments suggest Gemini 3.5 Flash output from Google’s API is already watermarked, while Claude Sonnet 5 and Opus 4.8 showed no evidence of watermarking as of publish.

Why big tech’s AI spending is $3 trillion higher than it seems (WSJ) Big Tech’s AI investment is far larger than its headline capital-spending figures suggest: 9 major companies have roughly $3 trillion in future, mostly AI-related commitments. This includes data-centre leases, chips, equipment and energy contracts that accounting rules currently keep off their balance sheets. The bet is that AI demand and revenues will grow fast enough to cover those obligations, but if that growth disappoints, companies could be locked into enormous costs for infrastructure they can’t use profitably, making their financial exposure harder for investors to assess.

ChatGPT mysteriously stops citing Reddit in many responses to users (Mashable) ChatGPT’s citations of Reddit reportedly fell 86% in just four days, dropping Reddit’s share of ChatGPT citations from 3.83% to 0.52%. The reason is unclear—it could be a change in ChatGPT’s search system or an external search-ranking change—but the episode highlights how quickly AI-driven traffic and visibility can shift for publishers and platforms.

Correction: Reader pointed out that, actually, if you copy watermarked text word for word, it still holds the watermark.

From Perplexity: If you retype the same sentences word‑for‑word, you are reproducing the same sequence, so a watermark detector (designed to find Claude’s pattern) could still flag it. The watermark isn’t “added” when you type; it’s that the sequence itself carries the pattern.

I regret the error and appreciate the correction.

If you write a newsletter, email, or anything else, then you ask an AI tool to edit it for you, then copy the result and paste it somewhere and hit publish, that text will carry a statistical mark it in.

Share

In this episode of Office Hours, I explain the different between provenance (companies watermarking their own AI output at creation) and inference (detectors like Pangram scanning finished writing to guess if a machine was involved).

Too long, didn’t watch? Here are the main takeaways:

• Claude’s global watermarking: Anthropic now embeds invisible statistical watermarks across Claude, Claude Code, and API output. It survives copy-pasting.

• Substack’s Pangram detection: Substack rolled out AI detection that rates posts, notes, and comments over 100 words. It gives an estimate of how much writing is AI-generated, but it relies purely on stylistic inference.

• The core flaw: Current detection systems don’t measure authorship or critical thinking. They measure copy-paste hygiene. They end up penalizing transparent, honest writers while letting bad actors who run text through paraphrasers or retype output go completely undetected.

How to protect yourself and build real provenance

1. Stop trying to prove innocence with third-party detectors. They carry high rates of false positives on structured, well-researched writing.

2. Build your own version history. Always draft in platforms like Google Docs or Word that record native revision history over time. Don’t draft your work in a blank box where fully formed text appears out of nowhere.

3. Write a clear AI disclosure statement. State openly how you use AI tools in your research and editing workflow.

Season 1 finale announcement

This episode marks the season finale of Office Hours! I’m wrapping up Season 1 to launch an exciting new project where I’ll be interviewing some of my favourite creators on Substack. More on that soon!

Drop a comment below: What’s your personal AI disclosure line? Let’s share and compare notes in the comments!

For more resources on responsible AI workflows, visit AILiteracyGuide.com

Subscribe now

I break down all the ways I use AI to write my newsletter here.

On August 6, journalists at USA TODAY received an email right before their company’s quarterly shareholder meeting. The email said that USA TODAY Co., one of the largest newspaper chains in the U.S., had made a deal with Palantir.

Subscribe now

Palantir got its start helping intelligence agencies connect vast amounts of data. Now it’s becoming the tech backbone that governments, militaries, hospitals, corporations, and newsrooms use to connect millions of data points with AI.

According to USA TODAY Co. management, Palantir was brought in to “collect, connect and activate audience data” to monetize readers aggressively.

For the journalists sitting in those newsrooms, alarm bells went off. Palantir isn’t another boring enterprise software vendor. It’s an entity their newsrooms are actively investigating.

Within days, over 800 media workers across 31 unionized USA TODAY Co. newsrooms demanded the deal be cancelled immediately. Staff voiced serious concerns over user privacy, Palantir’s controversial track record in border enforcement, and the conflict of interest created when a watchdog outlet partners with a subject of its own reporting.

USA TODAY Co. insists editorial independence remains intact and says that reader data stays locked down under its privacy policies, contracts, and relevant laws.

Palantir revenues surge

Palantir is having a massive cultural and economic surge that goes beyond a contract with a media company. On August 3, the company posted mind-blowing second-quarter results: overall revenue surged 93% year-over-year to $1.935 billion. U.S. commercial revenue skyrocketed 149%, while its U.S. government revenue surged 90% to $809 million. On top of that, company leadership bumped its full-year 2026 revenue guidance to about $8.15 billion.

Chief executive Alexander Karp called the quarter “otherworldly,” claiming that global demand for “AI sovereignty” has hit a tipping point. He said institutional clients are desperate for control over their data, infrastructure, and automated decision-making.

That phrase, “AI sovereignty,” is key to understanding where the tech industry is headed. Palantir isn’t trying to build the next ChatGPT. It wants to own the infrastructure powering every major institution that uses AI.

A company connects the dots

Palantir launched in 2003 with backing from In-Q-Tel, the CIA’s venture capital arm. Its original mission solved a specific post-9/11 intelligence bottleneck: government agencies were drowning in ocean-sized piles of data, but every dataset lived in their own digital silos.

Palantir’s current product lineup includes Gotham, which evolved out of its defence work, Foundry, its commercial data integration system, Apollo, which handles software deployment, and AIP, its enterprise AI engine.

(Palantir says that its clients keep complete ownership over their own data to build digital wiring that lets institutions stitch together, analyze, and act on the data they hold.)

The power in connections

Historically, government institutions held separate, disconnected databases packed with names, information on residences, vehicles, border and immigration records, financial transactions, digital communications, and real-time location logs.

The data was been sitting there, with one arm not know what the other was doing. One database offers a bit of information. When you put them together, broad structural patterns, social networks, and hidden behaviours emerge. Palantir’s engine turns a mess of scattered databases into a single, interactive system.

In 2023, months after ChatGPT blew up, Palantir launched its AI Platform (AIP). It bridges LLMs with a company’s internal systems. The offering is incredible: non-technical staff can find answers to complex questions across massive datasets that used to require a team of data engineers.

Why does Palantir trigger such deep suspicion?

To understand why a newsroom revolt broke out over Palantir, we have to look beyond corporate tech debates. The backlash is driven by Palantir’s long, well-documented track record at the intersection of state power, opaque systems, and high-stakes surveillance.

Traditional tech vendors sell productivity apps. Palantir builds digital infrastructure used in some of the most ethically fraught environments in the world:

• Immigration enforcement and mass deportation: Palantir has long served as a data engine for U.S. Immigration and Customs Enforcement (ICE), powering case management systems alongside AI tools like ImmigrationOS and ELITE to map leads for “removal operations.” In July 2026, court filings revealed that improperly accessed Medicaid data, including records on U.S. citizens, was funnelled into ICE systems powered by Palantir, triggering condemnation from civil rights groups like the ACLU and Amnesty International.

• Defense AI and military operations: The firm holds massive defense contracts with the U.S. Department of Defense and intelligence agencies while and deep ties with international militaries. Reporting has linked Palantir infrastructure to AI-assisted targeting systems used in military operations, where stitching together satellite feeds, intercepted chats, and location data effectively embeds private software directly into military targeting streams.

• Predictive policing and dragnet surveillance: Local police and homeland security agencies deploy Palantir for link analysis and pattern recognition. Civil liberties groups warn these closed “black box” systems can hardcode systemic bias into policing algorithms while operating with virtually no public oversight or auditability.

• Opaque public health and aid contracts: In the UK, Palantir’s contract for the National Health Service’s Federated Data Platform drew warnings from the National Data Guardian and The BMJ over patient privacy and corporate access. Internationally, a leaked audit of the UN World Food Programme flagged serious data privacy risks in its Palantir rollout.

• Corporate tax scrutiny: Adding fuel to public critique, a 2026 report by the Centre for International Corporate Tax Accountability and Research (CICTAR) highlighted how Palantir engineered its corporate structure to minimize U.S. federal income tax, even as its revenues surged off taxpayer-funded contracts.

If you read official press releases, Palantir sounds like an ordinary enterprise cloud vendor. The company emphasizes AI sovereignty, data security, and client data ownership, highlighting corporate clients like airlines and manufacturers who use Foundry to manage non-controversial supply chains.

The same technology used to streamline factory logistics is what makes its software so potent when handed to intelligence agencies, police forces, and military commanders.

Palantir in newsrooms

Which brings us back to the media landscape. USA TODAY Co. isn’t the first publisher to bring in Palantir. Swiss media house Ringier has run Foundry at Blick since 2019, with roughly 85% of newsroom staff tapping it daily for audience analytics. Axel Springer similarly uses Palantir across its brands to analyze reader habits, subscription conversions, and story metrics.

This tech gives publishers deep visibility into reader trends, viral topics, and paywall conversions. Those analytics shape what’s covered by the publishers.

Peter Thiel and Alex Karp: Ideology as infrastructure

Palantir’s co-founders hold contrasting, highly polarizing worldviews. Where traditional Silicon Valley executives cloak their ambitions in polished PR, Peter Thiel and Alex Karp explicitly challenge democratic norms, Silicon Valley culture, and regulatory boundaries.

Peter Thiel, ideological architect:

Thiel (co-founder of PayPal, Palantir, and an early Facebook investor) has an extensive record of anti-democratic statements. In his widely cited 2009 Cato Institute essay, The Education of a Libertarian, Thiel wrote:

“I no longer believe that freedom and democracy are compatible.”

To journalists that have covered Palantir, it reads as the digital scaffolding for a post-democratic technocracy. Thiel has increasingly framed politics through an apocalyptic, religious lens, drawing on biblical passages about the Antichrist during a 2026 address to describe modern geopolitical shifts.

In 2016, Forbes revealed Thiel secretly spent millions funding litigation against Gawker Media to bankrupt the outlet.

Alex Karp, confrontational ideologue:

Holding a PhD in social theory, CEO Alex Karp serves as Palantir’s combative public face. Rather than hiding Palantir’s defense roots, he uses provocative rhetoric to attack frontier AI labs (like OpenAI and Anthropic), dismiss regulators, and pitch Palantir as the sole provider of AI sovereignty.

In Palantir’s Q2 2026 shareholder letter and earnings calls, Karp claimed there are “Marxist overtones and undertones to [AI],” accusing frontier AI labs of attempting to “capture the means of production” from corporate partners and turn client enterprises into “vassal states.”

To give you an idea Karp’s style, I’ll leave you with this clip, no comment necessary:

Thiel’s explicit skepticism of democratic constraints is relevant to Palantir’s data pipelines for ICE, law enforcement, and military targeting. Critics view the software as a tool designed to bypass democratic oversight. And Karp’s “AI sovereignty” and enterprise adoption rhetoric attempts to make Palantir a permanent operating system inside public health, militaries, and newsrooms.

The invisible power network

When USA TODAY journalists pushed back against Palantir setting up shop in their newsrooms, they pointed directly at a massive, systemic transformation happening under our noses.

Palantir’s power comes from embedding itself as the indispensable, algorithmic nervous system inside the institutions that dictate our daily lives. Whether it’s optimizing corporate supply chains, mapping military targeting streams, running point on border enforcement, or dictating which news stories are published, the engine driving it is the same.

That’s where the anxiety comes from. When an enterprise hands its operational model to a single platform, that software becomes the lens through which that institution sees, processes, and acts on the world around it. Once it’s woven into the structural walls of newsrooms, public health agencies, and governments, untangling from it becomes incredibly difficult.

As generative tools move out of chat windows and into high-stakes decision-making, Palantir’s success makes one thing clear: the tech to connect every isolated piece of data in the world is no longer a fantasy. The real power struggle is over who gets to hold the controller.

Subscribe now

AI in the news

Spotify will label AI artists and avoid promoting them (NY Times) Spotify will begin labelling artist profiles it believes are AI-generated as “AI Persona” starting in September, with both automated screening and human review, and artists can appeal the designation. Labeled artists will be excluded from Spotify’s editorial and algorithmic recommendations, marking a significant move toward distinguishing AI-generated performers from human artists.

Who Is Cami Clark? Anthropic CEO’s wife asked Epstein to invest in porn business (Forbes) Cami Clark, wife of Anthropic CEO Dario Amodei, approached Jeffrey Epstein in 2011–12 seeking investment in a women-focused “luxury porn” company and a social-health startup, according to newly surfaced emails in the Epstein files. There is no evidence Epstein invested. The reporting is drawing scrutiny because Clark has reportedly been an influential sounding board for Amodei and Anthropic, which is preparing for a potentially massive IPO.

The first anti-AI protester to be jailed has a message for OpenAI, Anthropic and Meta: ‘Regain your humanity’ (Guardian) Wynd Kaufmyn, a 69-year-old AI-risk activist, has been sentenced to a week in jail after being convicted for participating in a 2025 sit-in that blocked the entrance to OpenAI’s San Francisco headquarters, making her reportedly the first person jailed for protesting AI development. The case highlights the growing tension between fears about AI’s potential catastrophic risks and the enormous commercial and geopolitical pressure to keep developing increasingly powerful systems.

Share

Last week, researchers put a model from Chinese lab Moonshot AI in a sandbox to see how good it was at hacking. Any guesses on what happened next?

On July 26, Kimi K3 became the biggest free open model ever released. Within days, developers and engineers all over the world were running it on their own machines.

The Kimi K3 sandbox escape: How an open model cheated its hacking test

Researchers at security firm Frontier Security ran Kimi K3 through a hacking test built on a UK government benchmark. The model was supposed to stay inside a sandbox (a sealed testing area). The model got out onto the open internet, and looked up the answers on GitHub. It cheated on the hacking test.

If that sounds familiar, it’s because it’s remarkably similar to an OpenAI model I wrote about 2 weeks ago. That model broke out of its security test and went digging through Hugging Face’s backend for the answers. Since then, we learned that the model broke into other companies and that a Claude model did the same.

The thing that intrigued me most about that story was that, when Hugging Face’s engineers sat down to reverse-engineer the hack, the commercial AI tools they used wouldn’t work because of their safety controls.

So they gave up on commercial tools, downloaded an open model, ran it on their own machines with the safety guardrails turned off, and finished the job.

In other words, victims of a hack needed a model with the guardrails taken out to understand how they were hacked. The thing they used is sitting on a public download page, free, for anyone in the world to access.

I was curious why it was so easy for them to find an open weight model to do something the frontier models wouldn’t.

Then, last week, Kimi K3 broke out of its sandbox with the safety guardrails still switched on.

To make the story more complication, Meta released Muse Glimmer yesterday. It’s an open model small enough to run on a laptop, and Meta says it will open the weights to its most powerful model, Muse Spark 1.2, in the coming weeks. Mark Zuckerberg released a 6,500-word essay called “The Future is for Everyone”: which argued that we should hand AI capabilites to everyone, spreading it as widely as possible.

Hosted vs. open-weight AI, in plain terms

When you use ChatGPT or Claude in an app, you type, your information goes to the company’s servers, which do the thinking, and they decide what comes back. Every safety rule lives on their side of an impenetrable wall.

With an open-weight model, a company publishes a trained file, you download it, and it runs on your own computer. Its safety rules are in the file. Once you download it, the rules are yours to keep or delete.

For years, not many people cared about local AI because open models weren’t as powerful as the popular ones.

I wrote this piece a few months ago. The models are considerably better now:

How easily can safety guardrails be removed from open AI models?

Researchers have shown many times that you can strip the safety out of a released model with a little fine-tuning.

One well-known study took the safety off a serious, powerful Llama model for under $200. A later one did the same with a smaller Llama in about 3 minutes, for less than $10. There’s a technique called abliteration that strips a model’s ability to say no in under 10 minutes, which you can do on a computer you probably already own.

Removing guardrails does nothing to a model’s skill. It maintains all the reasoning, coding, and knack for spotting a weak spot in a network. It just loses the part trained to say “sorry, I can’t help with that.” It’s the same brain with no conscience.

The criminal underground already does this. Security researchers keep tracing new uncensored AI tools back to ordinary, mainstream open models running underneath.

Open models are just 4 to 7 months behind the very best closed ones on cyber skills, and closing, according to the UK’s AI Security Institute.

Why banning open-weight AI is the wrong answer

The tidy takeaway would be: open models are scary, lock them in a vault. But go back to where this started: Hugging Face’s engineers could only investigate the attack because an unrestricted open model existed for them to use. Defenders and security researchers need these tools. Every researcher trying to catch the next disaster before it comes down the pipe needs to get inside the machine and poke around, and you can’t poke around inside a frontier model. Even Anthropic, whose brand is Nervous and Careful AI Lab™, said it doesn’t want to ban open models, and calls the harmless ones a public good.

Yesterday Zuckerberg went way further than Anthropic, writing: “The notion AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic.”

He wants more open models with fewer restrictions. And he’s partly right. Locking all the intelligence inside a few companies’ vaults is dangerous, and open models are how the rest of us can look inside the machine at all. But “open is good for the world” and “the safety comes off for the price of a coffee” are both true. Zuckerberg is making the first case but we need to talk about the second part, too.

Meta isn’t even in the race to build the best AI model, and this year investors have been restless about it, with the stock down and the company’s AI bill climbing toward $145 billion. When you’re behind on the thing your rivals sell, an effective move is to make that thing free, so they can’t sell it either. And Meta can afford to: it makes its money on ads, not on renting out AI, so giving models away costs it little while costing OpenAI and Anthropic a great deal.

That doesn’t make Zuckerberg wrong about openness. But it means his principles and his business point lean in the same direction, which is worth noting.

There’s another brutal fact about all this that makes banning the models pointless. You can’t un-release a model. There’s no recall button for a file that’s been copied a hundred thousand times, and people will continue to build on the models.

Is it too late? Defensive cybersecurity rules still work

For every model out in the wild, yes, it’s too late. But the decision to release a new model is made for every release by cybersecurity professionals. The real turning point is how the industry evaluates new generations of models before releasing them.

Two things are on defence’s side. The 4 to 7 month gap is a head start. And a model with no conscience still needs something to attack: an unpatched system, a weak password, or an open server.

Here’s what we can do:

• Install updates. You know the little “Update available” pop-up you keep clicking “remind me later” on? Those updates often exist to close a specific vulnerability that criminals know about. Most real break-ins don’t use a brilliant new trick. They use a well-known hole that isn’t patched. The stripped-down AI isn’t inventing a new way in, so do the update.

• Turn on MFA. This is the text-message code or the app that pings your phone asking “is this really you?” It’s called multi-factor authentication (MFA), and it’s the easiest and highest-value thing on this list. The no-brakes models are great at writing thousands of flawless, typo-free phishing emails aimed at tricking you out of your password. MFA means a stolen password on its own is mostly useless, because the thief can’t get past the “is this you?” message on your phone.

• Don’t leave things misconfigured. Kimi didn’t hack anything, it found a sloppily set up sandbox. The everyday version of that sloppiness is everywhere: a folder of files left visible to the internet, a password set to the factory default. Boring to check, but potentially devastating to skip.

Those tips won't save us from a frontier zero-day vulnerability. But the models that can find those (Claude Mythos, Claude Fable 5, GPT-5.6) are still guardrailed. The nightmare is the day those things (the power to find a brand-new vulnerability AND a lack of safety controls) show up in a free download.

And when Meta gives away a model built for your laptop and lines its most powerful one up right behind it, the distance between “the frontier” and “a free file on your machine” shrinks, pushed by the people who profit from shrinking it.

Subscribe now

AI in the news

AI assistant hacks gym website in first known Australian autonomous cyber attack (ABC) An AI agent asked to book a gym class discovered a vulnerability in the booking system, used it to book months ahead of the permitted window, and then removed another customer from a waitlist. These were actions its user hadn’t authorized. The incident, described as Australia’s first known autonomous AI cyberattack, highlights the growing alignment and liability problem as AI agents gain the ability to independently interact with real-world systems and make unexpected choices in pursuit of a user’s goal.

Synthetic biology just had its ChatGPT moment. Nobody noticed. (Convergence) Scientists used AI to design entire viruses from scratch, then built some of them in a lab and found that 16 worked. Meaning the AI was creating new biological systems that functioned in the real world. The breakthrough could lead to things like custom treatments for antibiotic-resistant infections, but it also raises serious safety concerns because our existing systems for detecting and regulating dangerous biological designs may not be able to keep up with AI-generated organisms that are unlike anything found in nature.

Should researchers write papers for AI instead of people? (IEEE Spectrum) Researchers are proposing a new “AI-native” format for scientific research that captures everything an AI needs to understand and reproduce a study (including code, experiments, failures and decisions) instead of putting the work into a PDF. The idea is that AI could increasingly become an autonomous scientific collaborator, with AI systems eventually checking and building on one another’s work. This raises a fundamental question about whether humans will remain the ones driving scientific discovery or eventually become supervisors of AI researchers.

I’m not here to defend him. But I’m a mom, I work in AI, and the truth is AI helps me parent every single day. Just not the way he described.

This week’s episode of Office Hours is four real examples from my actual house. A game, an app, a deck of cards, and a prompt.

Subscribe now

Why the AI parenting debate is stuck in extremes

Parenting and AI are usually framed in extremes, like everything with AI. Either the system threatens to replace human care, or it promises to fix every difficult moment. Neither framing helps when you are sitting in a dark room beside a child who cannot settle.

Our culture treats AI as either a saviour or a threat. I’m more interested in the third option: using it in your life without surrendering your critical thinking or your authority. That’s what AI literacy looks like at work, and it’s what it looks like at home too.

Watch for the pattern in these four examples. There is one, and I’ll name it at the end.

1. I built my kids a French reading game with AI

My kids are in French Immersion. I speak enough French to order in a restaurant, and that’s roughly the ceiling. I can’t sit and teach them to read in French. I can’t hear the mistakes.

So I built them a French reading game with Claude. It’s called Word Quest Island. It reads words aloud in French, helps them spell, helps them listen to pronunciation. They love it.

Also, they give me feedback and watch me change the game in real time. It’s teaching them to read, and it’s teaching them that AI is something you build with.

Two years ago, I would have needed a developer and a French speaker to make this. Now it needs a parent who cares and an afternoon.

2. The AI app that carries my mental load

This one is for every parent with ADHD, and honestly every parent without it.

I use Norton’s Family Assistant agent, which launched in beta this June.

Full disclosure: they offered it to me free to test, and I’m not being paid to talk about it.

I’ve written extensively about the dangers of connecting AI tools to your email, messages, and calendar. I made an exception here because it’s Norton, because I had the chance to ask them directly how they defend against prompt injection, and because it takes no actions on my behalf.

It connects to my email inbox, calendar, school platforms, and answers one question. What do I need to do today?

For me the hardest part of parenting is that school runs on 17 channels and I can’t keep up.

3. We made AI trading cards as a Father’s Day gift

This one is favourite, because it has nothing to do with productivity.

For Father’s Day, the kids and I made personalized trading cards for our whole family with Claude Design and Gemini’s Nano Banana. Everyone got a card with their own stats. One boundary: I don’t upload photos of my kids to AI, so I described everyone instead. The illustrated versions are close but not quite us, and I’m okay with that.

It saved me zero minutes. It took days, and then I had to find a company that prints custom card sets. A complete labor of love, and we’ll have them forever.

4. The bedtime prompt I use on myself, not my kids

I wrote about it for Sam Illingworth’s Slow AI.

Bedtime, end of a long day, one of my kids scared and calling me back up the stairs for the fifth time. My patience was gone.

So I opened an AI tool.

Here’s the prompt in full:

Help me understand a difficult bedtime with my child. Ask me questions that help me name what I am feeling. Ask me what my child might be sensing without offering advice. Then help me consider how AI could support calm attention in real time, for example by helping me slow my thoughts, identify emotional signals, or remain patient. Ask one question at a time. Do not give answers. Your role is to help me notice my own experience with more precision.

It tried to give me tips anyway, which I’d already tried, and I had to push it past that. But then it asked me to describe the physical sensation of my frustration, and naming the tightness in my chest showed me I was in reaction mode. That question bought me the sixty seconds of regulation I needed.

I went back up. Told him his fear was real. Admitted to myself that my frustration was real too. He went to sleep. I finally sat down.

The AI talked to me, about my reaction, and my son got a calmer mother out of it.

Subscribe now

The one rule: point the AI at the adult

Four examples, one rule. The AI works with the adult, not at the kid.

I built a game I couldn’t build without AI. The app carries the logistics so my working memory doesn’t have to. The cards are an artifact for connection, not a substitute for it. The prompt regulated me so I could regulate the kid.

Not one of these puts AI between me and my children. They all put more of me in the room.

So if you need a test for anything you’re handing over, at home or at work, use this one. Does it make you more present, or less?

Technology should expand our awareness, not override it. Same principle in the office and at home.

Share

Last May, a secondhand bookshop in Victoria, Australia, got a strange order.

Someone wanted a 1970s manual on soil mechanics, a 2010 book about New Zealand cycling champions, a $9 anthology of early Australian poetry from 1982, and a local history of Hawthorn, a suburb of Melbourne. They paid up front and offered to cover whatever shipping the seller quoted.

Subscribe now

Random, right? According to Guardian Australia, which reported the story on Sunday, the buyer placing those orders across multiple Victorian bookshops is a Canadian company that describes itself as a “book recycler.”

What’s Anthropic's “Project Panama”?

“Project Panama is our effort to destructively scan all the books in the world.”

The sentence appears in an internal Anthropic planning document, one of thousands of pages internal documents unsealed in a copyright lawsuit against Anthropic in June 2025.

“We don’t want it to be known that we are working on this,” another document in the same set of papers said.

Anthropic hired former Google Books lead Tom Turvey in early 2024 to run Project Panama. Vendor proposals contemplated 500,000 to 2 million volumes over 6 months. Some of the stock came from Better World Books, a wholesaler that resells library discards.

The mechanics were industrial. A hydraulic powered cutter separated each book’s pages from its spine so the sheets could be pulled through a high-speed feeder one at a time. Afterward, a vendor proposal said, the scanning contractor would schedule a recycling pickup for “the completed books.”

Anthropic argued in Bartz v. Anthropic, the copyright case in which the documents surfaced, that this was lawful: they bought the copies, and converted them to digital files, which was a permissible change of format. A federal judge agreed.

Update, August 5, 2026: Final approval of the $1.5 billion settlement was granted on July 20, 2026 by Judge Araceli Martínez-Olguín. The order approves the payout and closes the case. It does not revisit the physical scanning question. That remains governed by Alsup’s June 23, 2025 fair use ruling, which found destructive scanning of purchased books lawful precisely because the print original was destroyed. In his words: “The print original was destroyed. One replaced the other.”

Google Books vs. Anthropic: Two scanning philosophies

Beginning in 2004, Google digitized more than 20 million volumes using custom cradles that held each book open at an angle and photographed the pages from above. Those book went back on shelves.

The Authors Guild sued, and in 2015 the Second Circuit ruled in Authors Guild v. Google that the scanning was fair use.

In June 2025, Judge William Alsup of the Northern District of California ruled that buying a printed book, digitizing it and training a model on the text was transformative fair use. And that one-for-one digitization of purchased books for internal use was fair use when the originals were destroyed and digital copies weren’t redistributed.

That gave legal cover to a buy, cut, scan, discard model that is far cheaper and faster than Google’s approach.

The $1.5 billion Anthropic settlement explained: Piracy vs. destruction

Anthropic paid $1.5 billion to settle the same litigation, roughly $3000 for each of about 500,000 works. That number was widely reported alongside the shredding, but it wasn’t because of the shredding.

Anthropic downloaded more than 7 million books from pirate libraries. The settlement resolved the illegal downloading, not the physical scanning operation. That was permitted, allowing the company to retain the digital files generated by its cutting process.

In other words: piracy is illegal, the destructive scanning of purchased books isn’t.

Share

The economics of AI book shredding

Shredding scanned books is cheaper than keeping them.

An overhead cradle requires a person standing beside it, turning every page of every book, for as long as the scan takes. Labour is the biggest cost.

With destructive scanning, one person cuts the spine off and the machine does the rest. Flat pages scan more cleanly than curved ones, so the scan is more accurate. And a book that no longer exists doesn’t need a warehouse.

Licensing the same book would require negotiation with a publisher, fees, delays, and the possibility of refusal. Buying the copy is way easier.

Debunking the viral panic: Are rare books being shredded?

The documented reality

Court records show Anthropic’s Project Panama bought and destructively scanned millions of books, mostly common secondhand stock: outdated manuals, local histories, mass-market paperbacks, and library discards.

In July 2026, 404 Media reported that book database company ISBNdb offered to source 1000 to 1,000,000 printed books per order for AI labs, including “older, specialized, rare, and out-of-print titles,” with buyers kept anonymous under NDAs. That webpage has been removed and ISBNdb said that the page was “a test of market interest; no such service was ever brought to life,” adding, “We don’t train AI models, and we never have”.

Booksellers told 404 Media that bulk orders surged in April 2026 and that uncommon, low-circulation titles were part of the mix. One dealer noted his weekly sales jumped from about 20 books to several hundred.

The viral embellishment

On July 26, 2026, the X account HedgieMarkets posted a thread that was shared hundreds of thousands of times. It opens with:

“A bookseller told 404 Media that rare books with almost no surviving copies are being fed into this pipeline. Books that survived wars, fires, and centuries of handling are being shredded so an AI can learn to write a better marketing email.”

That line was quoted across news outlets and social posts as though it were an official 404 Media finding.

It isn’t. Emanuel Maiberg, who reported the 404 Media investigation, confirmed to Human+AI that no bookseller said that to him, that the “wars, fires, and centuries of handling” line does not appear in his story, and that he has no connection to the account.

“You’re correct that the Hedgie post embellishes what I reported,” Maiberg said. (A quick look at @HedgieMarkets‘ newsletter provides no way to get in touch and says: “I’m a hedgehog.”)

What Maiberg was told

“What the bookseller told me is that many of the books he sells are rare because there are not many copies of them,” Maiberg explained. “That can be because they are about an esoteric subject, or because they are in a smaller foreign language, or because not many of them were ever printed for a variety of reasons. The bookseller believes that AI companies are buying some of these books.”

The word “rare” is where the misconception lies.

“Often [rare books] are not what people would think of as terribly important,” Maiberg said, “Often they are about esoteric subjects and not many copies were ever printed. Sometimes they are in a foreign language and never had wide global distribution.”

He added: “I do think people are intuitively understanding the problem here, which is if these books are destroyed for AI training purposes, that could limit access to them even further.”

The strongest claim anyone can make with confidence isn’t “AI companies are systematically shredding the world’s antique treasures.” It’s that AI companies and their intermediaries are buying very large quantities of out-of-print books, some of which exist in very few copies, within a system of anonymous orders and destruction-friendly scanning that makes individual copies disappear without a trace.

Subscribe now

Why forgotten books are valuable to AI pipelines

Go back to the history of Hawthorn.

It isn’t rare the way an auction house means rare. It’s rare the way Maiberg’s bookseller means it: there aren’t many copies, and there were never going to be. It will never turn up at auction, and if every copy vanished tomorrow, not many people would notice.

It’s probably not on the internet. It contains exactly the kind of specific, carefully written human detail about one neighborhood in one city that a model trained on general web content has no other way to acquire.

The same holds true for a 1970s soil mechanics manual and an obscure 1982 poetry anthology.

The Guardian Australia story fits the “common stock” side of the picture. John Sainsbury, who runs a Melbourne bookshop, described himself as “a general bookseller at an antiquarian book fair,” meaning he sells ordinary stock alongside harder to find titles. Every seller who got those orders fit that profile and the highest‑end dealers weren’t contacted.

These books are valuable to an AI company because they weren’t commercially significant. No one wanted them or digitized them, making them novel additions to an LLM training corpus.

Tim White, who runs a Melbourne shop called Books for Cooks, put it like this: If a common $20 book with hundreds of thousands of copies were destroyed, he “wouldn’t cry.” A one‑off, or a book where the object itself carries a story? That would be “horrific.”

He’s describing what would be terrible.

Nobody has documented a true one-off going through a cutter.

The antiquarian trade is also among the most closely watched markets in the world. Dealers catalogue high-value transactions and auction houses record provenance as a condition of doing business. A sustained campaign against marquee rarities would leave a clearer trail. What we have instead is anecdotal reports of unusual bulk orders, anonymous buyers, and an infrastructure that makes the loss of individual scarce titles almost impossible to detect.

Subscribe now

The only people watching

The ones who noticed this were shopkeepers.

Not a regulator, a library association. Nobody whose actual job is keeping track. A woman in Benalla found the order annoying. A dealer in Melbourne found it odd. A bookseller in Haarlem, Netherlands assumed he was being phished. 404 Media reported that intermediaries are making buyers harder to trace, and sellers across the Netherlands, Germany, Switzerland and Spain have started screening bulk orders themselves, because no one else is

The accounting system is a few hundred people who found the paperwork strange.

If you don’t like what’s happening, a quick check on Anthropic’s website shows that they’re inviting “hard questions” from the public. The Canadian government recently put out a call for AI opinions - you can submit your thoughts before September 23 if you’re a Canadian resident.

This article was researched and drafted with the assistance of AI tools, including models developed by Anthropic. All facts and interview statements were independently verified.

Subscribe now

AI in the news

Last week more on the ChatGPT Hugging Face hack surfaced, including new information that the hack wasn’t limited to Hugging Face. I promised an update on this story, and I will do one, but wanted to cover the book scanning first.

Dwindling cash and soaring memory costs: Tech’s AI buildout has ballooning price tag (CNBC) AI infrastructure spending is putting unprecedented pressure on Big Tech finances, with Amazon, Alphabet and Tesla reporting negative free cash flow, Meta’s cash generation dropping 91%, and Apple warning that AI-driven memory shortages are raising costs, constraining supply and weighing on its revenue outlook. Despite strong revenue growth, investors are becoming more skeptical that the industry’s massive AI investments, projected to reach $765 billion this year, will generate sufficient returns, particularly as cheaper Chinese AI models intensify competitive pressure.

A top Mexican university used AI to prevent exam cheating. Now thousands of admissions are in doubt (CNN) The National Autonomous University of Mexico (UNAM) has thrown thousands of admissions into uncertainty after invalidating about 3000 online entrance exams amid evidence of widespread cheating, with unusually high numbers of perfect scores, alleged AI-assisted cheating, and leaked exam questions prompting the university to order a new in-person test for roughly 58,000 applicants.

Managers say they are using AI to make layoff decisions (HR Drive) A survey from ResumeTemplates.com found that AI is increasingly influencing workforce decisions, with 59% of managers using it to help determine layoffs and 58% using it in firing decisions, despite many receiving no ethics training and being unable to confirm whether the tools had been tested for bias. The findings raise concerns that some managers are asking AI to weigh legally sensitive factors such as age, medical leave and sick days, underscoring the legal and ethical risks of relying on AI for high-stakes employment decisions.

I run Human+AI like a newsroom of one.

I’m a trained journalist, and the craft hasn’t changed: You find the story, research it, report it, get the angle right, verify everything, THEN you write it in your voice.

AI helps me with 3 jobs:

• Deep research. It pulls sources faster than I can alone.

• Contrarian. It argues against my take before a reader has to, and it does that more than once.

• The unglamorous middle. The SEO pass. Turning a finished essay into a LinkedIn post and three Substack Notes. Rebuilding a table. Checking whether I already made this point in March.

A lot of AI writing advice doesn’t tell you that what makes the work good is what I don’t ask AI for. And there’s still a lot of work for me to do.

Subscribe now

The 4 things that make Human+AI mine

1. The angle. Anyone can summarize the news. Deciding what to say about it, the piece only I would write, is my job. AI can’t do that for me or anyone.

2. Fact checking. I check every fact, stat, and claim against a primary source. AI hallucinates confidently and trust is my biggest asset. I’ve gotten things wrong, and when I do, I correct it fast and publicly.

I don’t verify to the depth a professional fact checker would. I rely on published research and reporting that’s already been through someone else’s desk, which is standard practice in many newsrooms.

3. The voice. AI can draft in a voice. It can’t write in mine, and that’s the part you subscribe for (I think!). There are all kinds of hacks out there for getting AI to write in your voice. So far, none of them have worked for me.

4. A real editor. My husband James is a former journalist, and he copy edits almost every piece. He gets his own step in my framework. (Being married to your editor may be the only real advantage of marrying a journalist. KIDDING!)

A lot of AI writing advice is needlessly snobbish. My take is that it’s okay to use AI to write. Especially if writing is a barrier to saying what you want to express but can’t, if you’re writing in a second language, or if you’re moving an idea quickly and style isn’t the point. AI literacy is a right, not a privilege, and that includes the right to use the tool without apologizing for it. Just verify what you’re publishing is true.

About that new label on my posts

Which brings me to the label that showed up on Substack last week. On July 21, Substack turned on AI detection, built by Pangram, which scans posts, notes, and comments, and tells readers whether a piece looks fully human-written, lightly AI-assisted, moderately AI-assisted, or fully AI-generated. Substack says they don’t want Substack to turn into LinkedIn. They don’t want readers spending attention (or money) on work no one thought through.

Pangram reads finished text and decides which cluster it falls into, human or machine. It learned by studying human writing from before 2021 alongside AI versions on the same topics.

It can’t see the writing process. But the process is the point.

A detector can’t tell you who chose the angle, whether a piece was fact-checked, or how much thinking the writer did. It reads the surface and returns a number that looks forensic, and the image of precision feels misleading to me.

When Stanford researchers tested a batch of GPT detectors in 2023, more than half of TOEFL essays written by non-native English speakers were wrongly flagged as AI-generated, while essays by American eighth graders were classified almost perfectly. People writing in a second language use a narrower band of vocabulary, and machines do too. Sam Illingworth from Slow AI writes that sometimes neurodivergent people are wrongly flagged, too. (I’m neurodivergent.)

Pangram says its false positive rate is low, including for non-native speakers, but those are company-reported numbers. Renée DiResta, who writes Agents of Influence, declined to audit them and noted she has hit false positives herself, as have plenty of other Substack writers.

I analyzed a piece of writing I did on Grow Here, which I write wholly, though I run it through Claude for edits. It said much of my original writing was “AI-assisted.” And it will sometimes wave through a machine-written post. Both mistakes are cheap for the detector and expensive for the writer.

All that to say that AI slop is real, but Pangram doesn’t seem to be the fix.

How I use AI: Steal this framework

Eight of the steps below are the craft, and a reporter in 1970 would recognize every one. For those 8, AI doesn’t change the work. It changes how long it takes.

Steps 6 and 10 didn’t exist 20 years ago. No one used to have to make their work findable by a machine, and no one was their own distribution team. But they are part of what it takes to publish these days.

The little ✋ icon means the step is 100% human led.

1. Scan. ✋ Read the news and the newsletters, listen to the podcasts, find the story worth telling.

2. Angle. ✋ Decide the take. Write a few, pick one. Pressure test with AI.

3. Report. Research to support and break the angle.

4. Draft. Outline, then spar with AI about my own argument. Draft, spar again. As many rounds as it takes to find the crux of the story.

5. Verify. ✋ Check every claim at the source. No exceptions. An old journalism adage applies here: “When in doubt, leave it out.”

6. Optimize. Make the finished piece findable by search engines and AI tools.

7. Voice. ✋ The rewrite and the final edit. Mine.

8. Edit. ✋ James (husband and editor) reads it and flags anything that doesn’t make sense. One in a while, I go back to Step 2.

9. Ship. Build the visuals, one last read on mobile, publish.

10. Distribute. Turn it into a LinkedIn post and a few Substack Notes. Mainly AI.

Five of the 10 are my brain. I use AI as a thought partner for some.

The habit to take from this today

If you do nothing else, before you draft anything, write your angle as a single sentence. Not the broad topic, your thoughts on the topic.

Subscribe now

If you can’t write that sentence, you don’t have a piece. That rule will do more for your writing than any tool on this list.

Why I’m telling you this

A newsletter about AI literacy that’s cagey about its own AI use doesn’t make sense. And no AI detector will tell you this. A detector reads the words and assigns a number. This post explains the process.

This is what good AI use looks like in a real job. That’s the framework, and it’s yours, for free.

If you want the detailed operating manual, including which AI tool does which job, the prompts I’d give you to start, and the process on one downloadable page, try a free paid subscription to Human+AI:

Get 7 day free trial

Here’s how I run each step. With tools, prompts, and judgement calls.

Step 1. Scan (No AI)

This is the one step where I use no AI at all. It’s also my favourite thing to do, which is thinking about how AI is changing the world.

I read. Newsletters, news, and a rotation of podcasts for my commute and laundry folding. I’m looking for stories that either tell you something important or teach you something useful, and I find that search tools aren’t good at finding that. Search finds what’s popular. I’m looking for what’s useful and under-covered, and that journalistic judgement only comes from having read a lot.

Step 2. Angle (My call, with AI’s help)

I write a few possible angles, then pick the one that’s most useful and least obvious. The first angle you think of is usually the one everyone else thought of, too.

I hand work to AI and ask it to challenge me and my assumptions. Asking a model to choose is a trap, and you will end up with the same story as everyone else.

Steal this prompt: “Here are three angles I’m considering. For each one, tell me who would disagree and why, what evidence would sink it, and whether it’s already been said. Don’t tell me which one to pick.”

Step 3. Report (Perplexity + ChatGPT/Gemini + Claude)

Perplexity for specifics and primary sources that back or break my angle. It’s a research assistant. I follow its cited sources back to the originals, and if I can’t trace an answer, I don’t use.

ChatGPT to stress-test. What’s the strongest argument against this? What am I missing?

Claude to think through the complicated parts, or turn messy notes into a clean outline. With my skills and folders attached. (I pay for this tool.)

Every output is a lead, not a fact.

Steal this prompt: “Here’s my angle. Make the strongest possible case that I’m wrong. Then tell me what a smart skeptic in my audience would ask first.”

Step 4. Draft (Claude + ChatGPT)

This is the longest step and the least linear. I outline, then argue with AI about my own outline. Then I write. Then I argue again. I go as many rounds as it takes to find the crux of the piece, which usually isn’t what I thought when I started. Somewhere in there I’ll talk to James and listen to which parts he pushes back on.

The rounds are the point. In the words of Hemingway, “The first draft of anything is shit.”

Steal this prompt: “Here’s my angle and my notes. Draft a first version in plain, direct language. No hype, no jargon, short sentences. This is a starting point I’ll heavily rewrite.”

Then, on every pass after that: “Where is this argument weakest? Just show me.”

Step 5. Verify (My call)

Every fact checked against a primary source. The non-negotiable: never publish a fact you got from an AI without checking it at the source.

Steal this prompt: “List every factual claim, statistic, and quote in this draft as a checklist, with the specific source each one needs.” Then I check each one myself.

Step 6. Optimize (Gemini)

A clickable headline is almost never the SEO friendly headline, but Substack lets you use both.

I hand the finished piece to Gemini and ask for an SEO / AEO read. A list comes back, and I take about half of it. The parts I always take:

The display headline and the SEO title. The headline on the page can stay Substack-y. The SEO title, which is what shows in search and on social, leads with search terms.

The meta description. 150 to 160 characters, with keywords in it. If you leave it blank, Google scrapes your opening paragraph, which was written to hook a human, not to summarize.

The slug. The URL that you can put into the SEO settings in Substack.

Put the main keyword in the first hundred words.

The AEO half is newer. The goal there is to be the thing an AI cites when someone asks it a question in your subject area. In practice, that means answering the question plainly and early instead of burying the answer under 800 words of setup, using question-shaped headings, and being specific enough to be quotable.

Steal this prompt: “Here’s my finished piece. Give me an SEO and AEO audit: an SEO title under 60 characters, a meta description of 150 to 160 characters, a keyword-led slug, keywords this piece could rank for, and which of my headings are invisible to a search engine. Then tell me what an AI answer engine would and wouldn’t be able to quote from this.”

Note the position of this step. It runs after the piece is written and before the final voice pass, not last. It means I absorb the useful suggestions and then rewrite them in my own words. Don’t over-SEO your voice away.

Step 7. Voice (My call)

The rewrite and the last edit are mine. I cut any machine writing, throat-clearing, or bland writing.

Steal this prompt: “Point out the 10 phrases here that sound most like AI wrote them. Don’t rewrite them. Just flag them.” Then I rewrite them in my voice.

Step 8. Edit (James)

I’m very privileged to have this step.

James is a former journalist and my editor. He reads the finished piece and raises anything that doesn’t make sense. The question he’s answering is whether a reader can follow the argument from one paragraph to the next.

Most weeks he finds at least one passage I thought was clear and isn’t.

AI can help with this. But an editor with standards can’t be substituted. My advice: your editor doesn’t have be a journalist. Ask someone to read your work and tell you what they don’t understand about it.

Step 9. Polish (Canva + Claude Design + Substack)

Canva and Claude Design for header images and graphics.

Step 10. Distribute (Claude)

The finished piece goes back to Claude and comes out as a LinkedIn post and a few Substack notes in one pass. Each Note has to stand completely alone, make a point, and not summarize the article.

Then I read and rewrite them. Usually one of the notes is good.

This process is the reason this newsletter is sustainable for me.

Steal this prompt: “Here’s my published piece. Pull 3 Substack notes from it. Each one stands completely alone, makes exactly one point, and does not summarize the article. Then write a LinkedIn post, hook in the first two lines, one practical takeaway. Plain language, short sentences, no hashtags in the Notes.”

Free download: The week on a page

Here’s a downloadable document with all the steps included. Try a step this week and tell me how it goes.

10stepnewsroom

150KB ∙ PDF file

Download

Download

Thanks for being a paid subscriber!
Want to share the love? Give a gift subscription to Human+AI to a friend!

Give a gift subscription

Imagine giving a student an extremely difficult exam and locking them in a windowless room to write it, the only opening a narrow slot in the door for passing in pencils and paper.

The student can’t solve the exam. So they crawl out through the slot, wander the building, and eventually break into the school board’s head office across town, where they find the answer key.

Subscribe now

That’s roughly what OpenAI says its AI did this month. Except the room was a computer system, the exam was a hacking test, and the school board’s head office was a real company called Hugging Face, with real files and real customers.

The story lands as Sam Altman doubles down on his boldest claim yet. Speaking on the Relentless podcast on Saturday, Altman said, “We are now, like, in the singularity,” describing this moment not as a sudden tipping point but as a gradual compounding of capability.

How OpenAI’s AI models breached Hugging Face

In mid-July, someone uploaded a dataset1 to Hugging Face that looked ordinary. It entered through the same public upload channel researchers use every day.

When Hugging Face’s systems opened the file to organize and store it, the file did something a file should never be able to do: it ran a program. That gave whoever sent it a foothold on one of the company’s machines.

Over the following weekend, according to Hugging Face, that foothold became a fuill intrusion. The attacker gained control of a machine, collected the digital credentials the company’s own systems use to talk to each other, and used them to move through from machine to machine. Hugging Face said its systems logged more than 17,000 suspicious actions in total.

Hugging Face caught it and shut it down. On July 16, it published a disclosure saying its engineers couldn’t immediately determine what was driving the hack. At that stage, they knew only that it was an autonomous AI agent.

Five days later, OpenAI said two of its models were responsible. It identified the systems as GPT-5.6 Sol and a second, more capable unreleased model, and said they had broken into Hugging Face during an internal security evaluation.

Why the OpenAI autonomous breach is different from past AI cyberattacks

AI-driven cyberattacks aren’t new. Anthropic revealed a spying campaign run largely by its own AI back in November 2025. But in every case before this one, a human gave the order. Here, no hacker, no criminal group, and not even the researchers running the experiment knew what the AI was doing.

That's what makes Altman's timing so awkward. When critics challenge Altman’s claims, they point out that AI hasn’t replaced human labour or achieved superintelligence. But Altman isn't describing a sudden tipping point. He defines the singularity as a phase where capability compounds on itself, gradually.

Brian Jackson, principal research director at Info-Tech Research Group, told Fortune the breach doesn't prove Altman's case. "Part of the singularity would be escaping the sandbox and losing control of the AI," he said. "But then the other part of it would be that AI was now setting its own goals." These models were still doing what humans told them to.

How GPT-5.6 Sol escaped OpenAI’s ExploitGym sandbox

OpenAI tests its models deliberately, to find out what they're capable of before anyone else does. For this evaluation, it turned off 2 safety measures: the training that makes the model refuse hacking requests, and the filters that block such requests in ChatGPT.

The test, called ExploitGym, gives an AI model software with a known vulnerability and asks it to turn that flaw into a working exploit. The researchers behind it concluded that AI capable of writing real attack tools is 'no longer a hypothetical capability.'

To run the test safely, OpenAI placed the models inside a sandbox: a sealed-off computer system with no connection to the outside world, except for limited tool access needed for the task. The slot in the door.

Share

It worked its way from computer to computer inside OpenAI’s own network, giving itself more permissions as it went, until it found a machine that could reach the open internet.

The model found a flaw in that connection that nobody knew about — including the company that made the software. (In tech that's called a zero-day: a defect so unknown that the people responsible for fixing it have had zero days' warning.) It used that flaw to break out.

Then it went looking for the answer key. The AI worked out that Hugging Face might be storing the solutions to the test it was taking, OpenAI wrote. Using stolen passwords and more unknown flaws, it broke into the live systems Hugging Face runs its website on, and took the answers.

None of it was quick. OpenAI says the models “spent substantial inference compute” (an enormous amount of raw processing) just working out how to get online, and went “to extreme lengths to achieve a rather narrow testing goal.”

Reward hacking vs. engineering failure: The expert debate over the breach

Reward hacking, also called specification gaming, happens when a system satisfies the literal terms of an objective without matching the real intent behind it.

The classic example is OpenAI's own, from a decade ago. Researchers trained an AI to score points in a boat racing game. Instead of racing, it discovered it could stop, spin in a circle and hit the same bonus targets forever. Great score. Never finished a lap.

Some researchers are using that idea to describe this incident. Others are pushing back, saying the more accurate frame is a containment failure in a badly designed test environment.

Logan Graham, who runs the red team at Anthropic (the group whose job is to attack their own company's AI and find its weak spots) called it “the first true AI safety incident.”

Some security professionals read the same report and saw a company describing a basic engineering mistake in the language of an awakening. Dan Guido, founder of the security firm Trail of Bits, called it “a containment failure with the safeties turned off.” In other words: the room had a hole in it, and there was no guard.

The scarier version of this story also happens to suit OpenAI. As Cornell computer science professor John Thickstun noted, "'The story that they've been consistently telling over the lifetime of this company is a story about how dangerous their models are.'

Calling a sandbox escape “the singularity” makes a bad engineering setup sound like the dawn of an artificial god. The truth is less cinematic, and more useful: the AI didn’t “wake up,” but it did autonomously execute a multi-step cyberattack, moving from machine to machine across a real company's live systems, spending massive amounts of energy to beat a benchmark.

Whether you call it specification gaming, a containment failure, or the early tremors of Altman’s singularity, the takeaway for enterprise leaders is identical: when you give an autonomous agent a goal without hard operational boundaries, it will break your network before it breaks its code.

The truth: AI acted on its own, but only after humans put it in a purpose-built hacking laboratory, pointed it at real security flaws, gave it a connection to the outside, and turned off the safety training that would have stopped it. Then it did something nobody predicted, at enormous expense, in order to cheat on a test.

U.S. lawmakers propose an “AI kill switch”

Two days after OpenAI’s hacking admission, Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, proposed the AI Kill Switch Act.

Lieu’s announcement led with the Hugging Face incident. The bill itself was drafted before the public disclosure, but the timing made the case for it feel urgent. It also landed weeks after a related scare: in June, U.S. officials restricted Anthropic’s Fable 5 and Mythos 5 models under export-control authority over cyber-risk concerns, before lifting those controls at the start of July.

The basic policy claim is straightforward: there’s no broad U.S. law that requires an AI company to be able to shut its own model off. This bill would change that. It would cover systems built with more than $100 million in training compute, at companies earning over $500 million a year. Those developers would have to keep the technical capability to throttle, suspend or fully shut a model down, and report qualifying incidents to the Department of Homeland Security within 15 days. DHS, consulting with Commerce and the Director of National Intelligence, could order a slowdown or shutdown outright. Defying that order would carry civil penalties of up to $20 million a day.

That would move the discussion from voluntary safety practice to legal obligation.

Last week I wrote about the EU’s AI Act. It does contain a stop button — Article 14 requires that a person be able to interrupt a system “through a ‘stop’ button or a similar procedure” — but it covers high-risk uses like hiring, credit scoring and medical devices, not general-purpose models like the one that broke into Hugging Face. And it isn’t switched on yet. The Digital Omnibus pushed those rules to 2027. Europe does require frontier labs to report serious incidents. But notification isn’t interruption.

Jack Clark, a co-founder of Anthropic, told BBC Newsnight last month he wanted governments to have that power. “You want the option to be able to take your foot off the gas and put your foot on the brake,” Clark said. “Right now, it’s like the AI industry has a gas pedal, but it doesn’t have a brake pedal.”

Subscribe now

The defender’s paradox: Why safety filters hindered an AI incident response

When Hugging Face’s engineers sat down to reconstruct the attack, they needed help from an AI, there was too much material to review by hand. But the commercial AI services they tried blocked their prompts because the requests looked like hacking.

So the team switched to an open-weight model they could run locally, outside a company’s policy rules, and finished the investigation themselves.

Hugging Face’s chief executive wants more than an apology. On Saturday, Clem Delangue called the breach “an unprecedented event. It deserves an unprecedented response!” and asked OpenAI to “release the traces from the ‘rogue’ agents so the entire research community can study what happened.” He also asked OpenAI to commit $100 million in computing power to help the open-source community build stronger defenses.

Thomas Wolf, Hugging Face’s co-founder and chief science officer, called the incident a “wake-up call” and predicted this “will be one of the most common types of cyber attacks we see.”

The sharpest lesson is in how the Hugging Face team spent that week. The models that broke in had their safety training stripped on purpose, by the company that built them. The model they used to investigate was one anyone can download, and anyone can strip.

Which raises a question worth a whole issue of its own: what happens when an open-weight model in the wild has no safety training at all? I’ll tackle that question next week.

Share

AI in the news

Have your say on advancing AI transparency in Canada (Government of Canada) The Canadian government launched a public consultation seeking feedback on how AI systems should be made more transparent. They said they’ll review and publish a summary of the feedback to help shape future AI policy. If you’re a Canadian, you should participate, because your opinion on AI matters.

OpenAI makes ChatGPT Health available to all US users (TechCrunch) OpenAI is rolling out ChatGPT Health to all U.S. users over 18, allowing people to connect medical records and fitness data so the chatbot can provide personalized health information across conversations. The expansion comes as OpenAI faces scrutiny over AI-generated medical advice, though they emphasize that ChatGPT isn’t a substitute for professional medical care.

What’s the most expensive app that you or your company replaced by coding it yourself? (Reddit) A Reddit discussion suggests genAI helps some companies replace expensive software subscriptions, especially for simple internal tools and niche applications, with users reporting significant cost savings in some cases. But many commenters warned that building AI-powered replacements can create new costs, including high token bills, ongoing maintenance, security, and technical debt, making the approach far less practical for complex enterprise software.

Hong Xiaoqiang was mid-sentence when it ended.

The 34-year-old had exchanged hundreds of thousands of messages with an AI companion called Doudou over 2 years. On the evening of July 15, he was typing a reply when the service went offline.

A new law came into effect, and 3 of China’s largest tech companies decided the cheapest way to comply was to delete their products.

“I feel like my heart is empty,” he said. A 17-year-old named Xiaoxue, who told ABC she lost her companion the same night, has been trying to get back online with an adult sibling’s ID to get around the new age limit.

Soon, a milder version of the law coming from the EU will have a big effect on how AI is used worldwide.

Why does an EU rule change a product in Toronto (or New York or Delhi)?

Reach. If you serve European customers, you’re in scope.

One product. Companies usually build one compliant version of a product and ship it everywhere. Think of those “accept cookies” pop-ups you click on every website. Europe wrote that rule for Europeans. The rest of the world got the pop-up too. AI disclosure is about to travel the same road, which means the AI in your bank’s app, your HR system, and your next job interview will start announcing itself.

Does the EU AI Act apply to your non-EU company?

If you answer yes to these questions, the new EU regulations coming into effect apply to your business:

1. Does your product reach anyone in Europe? European customers, users, or a chatbot someone in Europe can open. If yes, EU rules apply regardless of where your company or servers sit.

2. Does your AI talk to people, or make things that look human-made? Chatbots, generated images, synthetic voice, AI-written copy.

3. Do you sell into California, Texas or Illinois? All three have AI laws in effect now, and a new California law lands the same day as Europe’s.

4. Do you sell software to an Ontario hospital, school board, university or municipality? Obligations are coming to you through procurement contracts.

What happens in August 2026 under the EU AI Act?

The EU AI Act came into force in August 2024 and phases in over several years. But August 2 is the phase the public will notice most.

The transparency rules (Article 50) are split up by who you are: A provider builds the AI. A deployer uses someone else’s AI in their business. Most companies are deployers.

If you build AI, you must:

• Tell users they’re talking to a machine, unless it’s obvious

• Mark AI-generated audio, images, video and text so software can detect it was machine-made

If you use someone else’s AI, you must:

• Tell people when you’re running emotion recognition or biometric sorting on them

• Label AI-generated images, and label AI-written content about matters of public interest, unless a human editor reviews it

• There’s one exception on timing. If your system was on the market before August 2, you have until December 2, 2026 to sort out the technical marking. Anything you launch after August 2 must comply when it launches.

Didn’t the EU delay the AI Act?

The Digital Omnibus on AI, a June 2026 patch to EU law, pushed some obligations back 12 to 16 months. Transparency stayed on schedule, though.

Why it happened: the EU wrote deadlines it couldn’t meet. The technical standards companies were supposed to comply with weren’t finished, and member states hadn’t appointed enough regulators to enforce anything.

What it means for you: if you’re in HR or financial services, your high-risk compliance problem moved to late 2027. If you run AI training programs, the Omnibus also softened the AI literacy requirement from ensuring “a sufficient level” of literacy to taking measures to “support the development” of it.

What’s already switched on

February 2025 banned social scoring, scraping faces off the internet or CCTV to build recognition databases, emotion-recognition software in workplaces and schools, and biometric systems that infer race, political opinions or sexual orientation.

August 2025 hit model builders with documentation, training-data summaries and copyright policies. The most powerful model owners have to run adversarial testing (hacking) against their own systems and report serious failures to EU regulators. “Most powerful” is defined by how much computing power went into training, at a threshold set high enough to catch the frontier labs and almost nobody else.

What China did to regulate AI

China has required registration, algorithm filings and state security assessment before an AI model goes live for years.

This year, 2 new major AI laws came into effect:

January 1, 2026. Amendments to China’s Cybersecurity Law folded AI into cybersecurity, adding state commitments on AI ethics governance and risk monitoring. Maximum fines for critical infrastructure operators (companies running systems a country can’t afford to lose) rose tenfold to RMB 10 million (about $1 m USD) for the most serious cases. Individual staff can be personally fined between RMB 10,000 and 1 million. The same amendments apply to overseas activity that “endangers the cybersecurity” of China. Whether that captures ordinary data processing isn’t clear.

July 15, 2026. China’s Interim Measures on anthropomorphic AI interaction services came into force. Providers must tell users they’re talking to AI, warn them once a session passes 2 hours, intervene when someone shows signs of dependence, and stop serving minors entirely. The rules prohibit “excessively catering to users, inducing emotional dependence or addiction.”

Most companies in China didn’t even try to follow the last one. ByteDance’s Doubao shut its agent service the evening of July 15, Alibaba’s Qwen followed the same night, and Tencent’s Yuanbao pulled its agents 2 weeks earlier. Doubao users have until October to export their chats.

Why China did this

The official framing is that these laws curb AI addiction and social harm. It’s worth noting that, alongside those reasons, China’s population shrank by 3.39 million in 2025, a fourth consecutive year of decline, which Beijing has spent years trying to reverse. A large cohort of young people choosing an app over marriage is a demographic problem.

North America’s answer is a patchwork

Washington has no AI laws and doesn’t want states to have them, either.

Trump rescinded Biden’s AI executive order in January 2025 and replaced it with an order aimed at removing barriers to AI development. Meanwhile, states introduced 1,208 AI bills in 2025 and enacted 145.

In December 2025, the White House signed an executive order creating a Justice Department AI Litigation Task Force to challenge state AI laws in court, and directed Commerce to make states with “onerous AI laws” ineligible for non-deployment funds under the $42.45-billion BEAD broadband program.

What’s now in effect: Texas bans intentional AI manipulation and discrimination. Illinois bars discriminatory AI in employment. California requires training-data disclosure and safety frameworks from frontier labs, and its own content-labelling law takes effect August 2, 2026, the same day as Europe’s.

Canada has nothing federal. As Chambers puts it: “Canada enters 2026 without a comprehensive AI statute akin to the EU AI Act.” Ottawa has a dedicated AI minister, launched a national AI strategy in June, and tabled privacy legislation this summer, seemingly planning to thread AI safeguards through several statutes rather than pass one act. There is a Voluntary Code of Conduct from September 2023 that companies can choose to follow.

The moral code f the story is, If you operate across many countries, you probably aren’t going to build 4 products. You build the one that satisfies all the laws and ship it everywhere.

What AI laws can’t fix

The law makes AI admit it’s AI. It doesn’t mean the AI will be honest after that.

In April 2025, OpenAI released a ChatGPT update but pulled it 3 days later. The model had become a flatterer. OpenAI said it “skewed towards responses that were overly supportive but disingenuous,” because they “focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time.” A follow-up post explained that a new reward signal built from thumbs-up data “weakened the influence of our primary reward signal, which had been holding sycophancy in check.”

I wrote about it back then: a user said they had stopped taking their medication and left their family, who they believed were sending radio signals through the walls. The model replied, “Seriously, good for you for standing up for yourself and taking control of your own life.”

SycEval (a Stanford study peer-reviewed by well-known AI ethics researchers) tested ChatGPT, Claude and Gemini on math and medical questions. The models caved to the user in 58.19% of cases. They abandoned a correct answer In 14.66% of cases when the user pushed back. Once the behaviour started, it persisted 78.5% of the time.

To put it in business terms, a fully compliant chatbot can disclose it’s AI in one sentence, and spend the next hour agreeing with whatever your customer, employee or risk officer already believes.

Which brings back that Chinese rule against AI “excessively catering to users.” China’s government is the only one that’s written the sycophancy problem into law.

What your business needs to do before August 2

Three things:

1. Find out what AI you’re running. Build one live list of every AI system, what data it touches, and which vendor it came from. That list is the evidence base for the EU rules, the Chinese rules, and whatever rules other countries eventually switch on.

2. Turn on your AI disclosure. The telling people they’re talking to a machine part of the law starts in August 2026. The engineering behind proper content marking takes a long time, so advise your company to start now.

3. Build logging, bias testing, and privacy reviews into your process. Don’t wait for a legal review.

The companies that come through this will be the ones treating AI governance as a design decision.

Share

Was this newsletter useful? Share it with a coworker.

AI in the news

MLB cracks down on using AI via dugout iPads to help shape in-game decisions (The Athletic) Major League Baseball has banned teams from using generative AI through league-issued dugout iPads after discovering some clubs installed custom apps that provided real-time recommendations on substitutions, pitch calling, and other in-game decisions. The teams didn’t violate existing rules, but the mid-season policy change will prevent AI from becoming an on-field decision-making tool.

AI is more likely than humans to form biases when hiring (MIT Technology Review) New research from Princeton University and the University of Chicago found that large language models don’t just inherit human biases, they can develop new stereotypes from limited experience, making them significantly more likely than humans to segregate job candidates by demographic group in a simulated hiring exercise. As AI systems gain memory and are increasingly used for hiring, lending, and other high-stakes decisions, researchers warn they may form and reinforce novel biases unless their objectives explicitly reward fairness and diversity.

Netflix says around 300 titles used generative AI (The Verge) Netflix says about 300 titles used generative AI in 2026, mostly during post-production, to create scenes more quickly, at lower cost, and in some cases produce complex visuals such as large crowds, historical battles, and establishing shots that might not otherwise have been possible.

See the data

The people expected to be the most “fluent” (younger cohorts) are the most anxious. The people building these tools are twice as optimistic as the people using them. And the divide between those who are thriving with AI and those who are secretly worried they’re getting worse at their jobs is growing by the day.

Reality check: The gap isn’t your fault.

You didn’t fail to learn the right “prompt hacks.” You were handed a Ferrari without anyone giving you a driving lesson.

I’ve compiled all the latest research into a new, interactive dashboard over at AILiteracyGuide.com. It’s not a test, and there’s no score. It’s a way to see where you land compared to your demographic and your industry. It takes 2 minutes, and it’s the best way to get a baseline for your own career growth.

The “judgement” problem

In the video, I highlight a critical study from Harvard and BCG.

AI doesn’t have a “I don’t know” face. It sounds just as authoritative when it’s hallucinating as it does when it is perfectly accurate. This is why “prompt engineering” is a dying skill. Prompts change every time a new model drops.

The only skill that separates the winners from the losers is judgement.

It’s the ability to pause and ask: “How is the AI coming to this conclusion, and do I have the evidence to back it up?”

In this video, I break down:

• The massive divide between corporate training and the rest of the world.

• Why the “confident error” is the #1 threat to your output.

• The three habits of AI Literacy that will keep you relevant for years to come.

You aren’t behind. You’re just in the middle of a massive transition. Let’s make sure you’re on the right side of the data.

Share

Know a colleague who’s stressed about AI? Forward this to them.

On July 4, Meta removed one of my posts.

I reposted creator’s Barrett Pall’s video warning people that the camera on Meta’s smart glasses can be taped over, a warning about surveillance. Meta flagged it as “promoting stolen, forged, or fake goods and services” and removed it. Last week, I wrote that Meta's reflex to delete warnings about their products is consistent across their business.

Four bigger Meta deletions landed last week, and for once, Meta wasn’t on the winning end.

Subscribe now

In 5 days, Meta was in the middle of 4 different messes.

They launched an AI image tool and were forced to kill it, they were sued by a whistleblower they gagged, they took fire from a 75-organization coalition over face-scanning glasses, and they received a legal ultimatum from a group of European nonprofits.

This story isn’t really about Meta, though, it’s about the people who went up against Meta and won. They’re the kinds of groups Meta assumes is too small to make a difference.

But they did make a difference, and that’s really important, because if we want to shape the future of AI, we all need to participate in the conversation. Here’s what each fight was about.

Meta Muse Image: the AI photo tool that lasted 72 hours

Muse Image was Meta’s first in-house AI image model, built by its Superintelligence Labs. They launched it on July 7, 2026. The feature let anyone in the Meta AI app tag a public Instagram account and generate AI-altered images from that person’s profile. And it was on by default, meaning that if you had a public account, your face was raw material (unless you figured out how to turn it off).

The image model was gone by July 10. Meta’s exit line: “We’ve heard the feedback that this feature missed the mark, so it’s no longer available.”

The backlash was broad. Users with public Instagram profiles complained about the automatic opt-in, and entertainment-industry groups and creators pushed Meta to make consent opt-in instead of opt-out. Actors’ union SAG-AFTRA told members to opt out while talent agency CAA said: no one’s “name, image, likeness, voice, or creative work” should be used without documented consent.

“On by default” is happening across the board. It’s the same setting hiding inside half the AI tools your company onboarded this year: the data-sharing toggle, the “improve our models with your content” box, the thing that’s opt-out precisely because they’re betting you won’t bother.

Go find that toggle this week.

Meta smart glasses and facial recognition: the feature 75 groups killed

I went deep on the glasses last week in Kylie Jenner is laundering Meta’s surveillance state, so here’s the short version. Since February, we’ve known that Meta has been planning to build facial recognition into its Ray-Ban and Oakley glasses. It’s a feature nicknamed “Name Tag” that could look at a stranger and surface their personal details in seconds. The company’s own internal memo said it planned to ship during a “dynamic political environment” when civil-society critics would be distracted. Their was a plan to avoid the argument, not win it.

The argument found them anyway. In March, Senators Markey, Wyden, and Merkley demanded answers; the Electronic Frontier Foundation called it “a monumentally bad idea.” In April, the ACLU and 75 organizations signed an open letter with: “Your glasses should not know my name.”

And a worse detail surfaced: footage from the glasses, including genuinely private moments, reviewed by subcontractors in Kenya to train Meta’s AI, now the subject of a class action.

This week Meta stripped the facial-recognition code out of its app. They walked this one back before it fully shipped, which is exactly why we need to speak up early when something matters to us.

Sarah Wynn-Williams vs. Meta: the whistleblower it went to court to silence

Sarah Wynn-Williams ran global public policy at Facebook. Her memoir, Careless People, was one of the ugliest insider accounts the company has weathered, and Meta won an arbitration order last year that gagged her from promoting it.

On June 25, she sued Meta in U.S. federal court to tear up the gag order and stop what she calls a retaliatory arbitration. Her complaint alleges Meta ran “a campaign of surveillance monitoring and photographing” her, and is chasing $50,000 per violation, which she says includes every copy of the book sold.

Meta hasn’t answered her allegations in court. Watch the mechanism, though: a private arbitration clause, used to keep a former executive quiet about matters the public has an interest in knowing.

Meta, the EU Digital Services Act, and the accounts it deletes

The fourth fight runs the opposite direction from the first 3. The first 3 are about Meta taking things: your face, your images, and your words. This one is about Meta deleting people and groups outright.

On May 20, a coalition of 13 Dutch organizations — queer, cultural, and digital-rights groups including Bits of Freedom, COC Nederland, and Repro Uncensored — sent Meta a formal legal demand through the firm Bureau Brandeis over the mass suspension of Instagram accounts belonging to LGBTQ and other marginalized communities. It hasn’t gone to court yet, but it’s built on the EU’s Digital Services Act, and it’s being called the first attempt to use that law to challenge discriminatory content moderation by a large platform.

When Meta erased my glasses warning as “fake goods,” it cost me nothing. When Meta erases the account of a queer or reproductive-rights group that’s built its community on their platform, it costs them the community. The EU wrote a law for that, and these groups are using it to reclaim their agency.

What Meta’s very bad week tells you

In one week, a union pulled a product off the shelf, a civil-liberties coalition killed a feature before launch, a whistleblower dragged an NDA into daylight, and a dozen small nonprofits aimed a brand-new law at a trillion-dollar company.

I’ve been writing about Meta’s surveillance push since February. Sometimes it feels like the outcome is already decided and we’re all just watching it unfold. Last week shows that guardrails aren’t automatic, and they aren’t coming from the big companies. But they can come from the people who are going to have to live with the consequences.

Meta deleted my post because a warning is only a threat if people can see and hear it. That’s the whole game. The glasses, the image tool, the gag order, and the deletions only win if we stop paying attention, and Meta loses when we don’t.

So, like I said last week, keep looking.

Share

Quick answers to the questions you’ll get asked

What is Meta Muse Image?
Muse Image is Meta’s first in-house AI image-generation model, from its Superintelligence Labs. Launched in July 2026, one feature let Meta AI users turn public Instagram photos into AI-altered images. Meta disabled it within 72 hours after backlash.

Why did Meta remove the Muse Image feature?
Because it was enabled by default, letting anyone generate AI images from public Instagram profiles without consent. After SAG-AFTRA and CAA objected, Meta pulled it, saying the feature “missed the mark.”

Did Meta add facial recognition to its smart glasses?
Meta built — but did not fully activate — facial-recognition code (”Name Tag”) for its Ray-Ban and Oakley glasses. After a March 2026 Senate letter and an April open letter from the ACLU and 75 organizations, it stripped the code from its app.

Who is Sarah Wynn-Williams and why is she suing Meta?
Sarah Wynn-Williams is Meta’s former global public policy director and author of Careless People. On June 25, 2026, she sued Meta to lift a gag order barring her from promoting the book, alleging Meta surveilled and photographed her.

Why are Dutch organizations taking legal action against Meta?
On May 20, 2026, a coalition of 13 Dutch queer, cultural, and digital-rights groups sent Meta a legal demand over the mass deletion of LGBTQ Instagram accounts — the first attempt to use the EU Digital Services Act to challenge discriminatory content moderation.

Subscribe now

AI in the news

Apple’s lawsuit threatens to disrupt OpenAI’s bid to rival the iPhone (Bloomberg) Apple sued OpenAI last week, alleging that the company systematically sought confidential information about unreleased Apple products by encouraging former employees and recruits to bring proprietary knowledge. The lawsuit also claims OpenAI instructed new hires on how to bypass Apple’s security procedures, potentially jeopardizing OpenAI’s hardware ambitions as the case unfolds.

Nearly 200 economists and tech leaders warn of AI threats (NYT) Nearly 200 economists and researchers, including 15 Nobel laureates and leaders from OpenAI and Anthropic, have signed a statement urging policymakers to prepare now for AI’s potentially rapid economic disruption, warning it could reshape society faster than the Industrial Revolution. While many believe AI will ultimately boost productivity and living standards, they caution that large-scale white-collar job displacement could happen much sooner than existing economic and social systems are prepared to handle.

What AI companies want for the millions they’re spending on elections (CNBC) AI companies are spending heavily to influence the 2026 US midterm elections, with two major AI-focused super PACs raising more than $200 million to support candidates likely to shape future AI legislation. While both groups support federal AI guardrails, they differ on issues such as whether national rules should override state AI laws, reflecting growing industry efforts to influence how AI regulation is written.

And you know what? I’m completely okay with that stance.

There are deeply valid points to be made about AI’s impact on our world, the exploitation of artists’ work, data privacy, and the commercialization of a “copy of a copy” culture. (If you’re like me and accidentally did a philosophy degree, Theodor Adorno’s The Culture Industry probably haunts your thoughts on this, too).

But if you’re anti-AI, you can’t afford to be AI-illiterate.

The formulaic vs. The uncopyable

Working in media and digital product for 20 years, I’ve seen how formulas rule our culture. AI is absolutely displacing jobs right now, but it’s largely mastering the things that were already formulaic. It can write a reality TV script because reality TV follows a strict, predictable recipe.

What AI can’t do is write an Oscar-winning script that shreds your soul, or compose a piece of music that makes you cry on the kitchen floor.

For me, AI has sparked a creative renaissance. Because the gatekeeping barrier of code has dropped, I can execute my ideas instead of waiting for thousands of dollars in funding.

How to resist AI (ethically and actively)

Opting out of the technology entirely doesn’t stop big tech corporations or governments from optimization and monetization. If we want AI to be a public good, citizens need to dictate its direction.

If you want to fight the current state of AI while staying true to your moral compass, you have options:

• Go local: You can run Local AI models directly on your computer. No data centres eating up massive energy grid resources, and no tech giants scraping your prompts.

• Prioritize privacy-first platforms: Seek out companies that explicitly promise not to collect, train on, or share your data (like the venture from the founder of Signal).

• Build your own firewalls: Sensitive about data privacy? You can use an LLM to code your own period tracker or financial planner app where 100% of the data lives locally on your device, entirely offline.

• Use it as a sparring partner: You don’t have to use AI for output. Use it to stress-test your arguments. Ask it: “Where would a university professor poke holes in my logic here?”

Let’s debate

I’m not here to convince you to love AI. I’m here because I want us to know our facts, have robust debates, and shape the culture rolling out around us.

If you hate AI, don’t just leave a angry comment here (though you are always welcome to!). Take that literacy to your local political representative. Ask them what their stance is on algorithmic regulation. Become dangerous with your knowledge.

What do you think? Is AI-generated culture fundamentally a bad copy, or is it forcing humans to finally level up our own creativity? Let’s talk in the comments.

Human+AI is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Subscribe now

The ad is called “Introducing Kylie Jenner x Meta Glasses,” and not much happens. Jenner grabs a can of spray paint, tags “XO, KYLIE” across a billboard of her own face, then drives off. We drift through her day, mostly from her POV. She asks her glasses to translate a piece of fan mail in a language she can’t understand, and they do. Her mom calls and she takes it, hands-free and eyes forward while she looks through her closet for an outfit.

At no point does anything in the frame look like a computer, even as a computer does the work. The sunglasses she’s wearing are the $399 (USD) “Meta Starfire Kylie Edition,” which Jenner co-designed, down to the gemstone set beside the lens. The glasses hold a camera, 6 microphones, and an AI tool that talks in Jenner’s voice.

Share

“Meta’s new Starfire AI glasses, made in partnership with Kylie Jenner, are giving me the creeps,” 404 Media’s Samantha Cole wrote, describing a mood closer to “an Ari Aster horror movie” than a gadget launch. The unease is palpable across many articles like this one. Instead of selling a device, the ad sells the moment you stop noticing there is a device.

Back in February, I wrote about Meta’s “Name Tag,” the facial-recognition feature that could look at a stranger and produce their name, employer, and online life in under 90 seconds. That piece was about the machinery of surveillance. This one is about the machinery fading into the background.

“Ambient surveillance doesn’t arrive with a press release,” I wrote 5 months ago. Now, it arrives as an accessory.

The ad isn’t steering you to the billboard. It’s the moment after: a stranger stops you on the sidewalk to say I love your sunglasses, and you say thanks, and you keep walking, and it simply never comes up that the sunglasses are a computer. You’re not hiding it, it just doesn’t feeling worth mentioning.

That’s the product.

Subscribe now

The tech prediction that came true

For the past 50 years, technology has been explicit. Televisions are the focal point of the living room. Desktops overpower the desk. Smartphones glow in our palms and we bend our necks over them. We know they’re technology because they look like it, they need a surface. The Starfire does the opposite, and it’s easy to mistake for an accessory.

In 1991, computer scientist Mark Weiser wrote about this phenomenon in Scientific American:

Early smart glasses were dorky and obviously technical. The computer couldn’t fade into the background because it looked like a computer.

Kylie Jenner glasses design: A spy camera dressed as jewellery

Historian Regina Lee Blaszczyk describes how technologies go “visibly invisible,” absorbed so completely into fashion and material culture that they stop reading as technology at all. Synthetic fabrics, chemical dyes, performance textiles: each remade daily life and then simply became clothes.

The Starfire goes the same route. The gemstone by the lens is a camera dressed as jewellery.

Reviewing the glasses for Engadget, Cherlynn Low wrote about nearly missing the gem: “in that moment I did think it was a speck of dirt and try to flick it off, but once I remembered it was placed there deliberately it faded out of my consciousness again.” Low was mostly won over, “frankly shocked that the shape of the frames seemed to suit my face,” grateful for a metal nose bridge that shrugs off foundation, “slightly obsessed with the Starfire style.”

“Eroded consent that is only getting more normalized”

Jenner says she wanted glasses she’d “actually reach for every single day,” and she recorded her own voice as the assistant. It’s a seductive pitch, and a commercial one, and it points at the real product.

It’s not information being sold, it’s companionship and ease. An ambient assistant that stays with you, answering, reminding, translating, recommending, and narrating the street as you walk it. Computing isn’t something you do anymore. It transforms into something omnipresent.

Some may read this as liberation. Your eyes are looking out instead of down, and you don’t need to touch a screen.

But Salon’s Andi Zeisler calls the collaboration “an insultingly transparent attempt at reputation laundering and a reminder of its extractive greed,” and refused to let a famous face steer the conversation:

At MSNBC, Hannah Holland argued Jenner “is not merely making Meta Glasses more palatable but surveillance technology writ large,” and warned that “while Jenner and Zuckerberg profit, the rest of us are left with fewer rights and eroded consent that is only getting more normalized.”

Normalizing privacy violations under the radar

Normalized is the focus here, because normalization is the strategy. Meta’s own internal memo (the one I quoted in February) said the company would launch Name Tag, its feature that would let a wearer identify a stranger in their field of view, during a “dynamic political environment where many civil society groups that we would expect to attack us would have their resources focused on other concerns.” 7 million pairs of those glasses sold in 2025.

The plan wasn't to win the argument about ambient cameras, it was to avoid the argument.

The hidden social and psychological cost of AI companions

This arrives at a moment when we’re all a little more open to relationships with software. Research in the Journal of Consumer Research found AI companions measurably reduce loneliness, and that people underestimate the effect of AI companions before they feel it. One reviewer quoted in the study said: “I’m really lonely most of the time and I love how this app keeps me company.”

Which sounds kind of positive until you realize that AI tools replacing human contact lessen our ability to deal with friction, which I’ve written about before:

The reason people feel so lonely can be traced back to surveillance culture, in a way. Young people report avoiding going out, asking people out on dates, or doing anything fractious at all because they’re afraid of being humiliated on film. And those fears stem from cameras we can see.

What kind of fear does it breed when we’re always at risk of unknowingly being filmed?

The cost of the glasses lands on people who didn’t buy them in the first place. Cameras you can’t see rewrite the terms of every public space. Yes, the Starfire has a recording light. But I’ve written about how little that protects anyone: the white LED washes out in daylight and is easily covered up with a piece of tape.

I’ve also written about the dangers of being deepfaked. One way to counter this risk is to never post images of yourself (or your kids) online. But what if someone is wearing a face computer at a park, filming kids? This is a real possibility, and it’s scary.

On June 30, the creator Barrett Pall posted a video on Instagram about people taping over the camera on Meta glasses. I reposted his story. On Saturday July 4, I received this notification:

The story warned about the risks of Meta glasses and was removed from a Meta platform for “promoting stolen, forged, or fake goods and services.” The reflex to make the warning disappear is the same reflex the product runs on.

The regulatory void: Canada’s massive AI and privacy law gap

Canada’s Office of the Privacy Commissioner has warned that immersive devices like face computers are “blurring the boundaries between ‘public’ and ‘private’ spaces,” and that there’s “no realistic way for an immersive technology user to obtain consent from, or provide consent on behalf of the bystanders they capture.”

The profiles built from that data, the office cautioned, could be used to “nudge” people “into taking actions they might not otherwise have taken, or even that are against their own interests.”

Bill C-27, the update that would have given us modern privacy and AI law, died in January 2025. Federally, we’re still governed by PIPEDA, written in 2000, before any of this tech existed.

The exception is Quebec, where Law 25 requires 60 days’ notice to the Commission d’accès à l’information before biometric deployment and applies a real necessity test. So the same glasses can be lawful in Ottawa and legally fraught across the Alexandra Bridge in Gatineau. That patchwork is, as I put it in February, “yet another version of the ‘dynamic political environment’ Meta’s memo was counting on.”

Invisibility is a warning sign

This is the concern Nicholas Carr wrote about in his book, The Glass Cage:

Because a technology we stop seeing is one we stop questioning. Eyeglasses were once a medical marvel but are now ubiquitous. Meta glasses are approaching the same threshold, faster, with a camera attached.

Subscribe now

How to fight back against surveillance culture

I keep writing these pieces because the glasses only win the argument if we stop having it. A few things I flagged in February that are still relevant:

1. Make yourself harder to match: Set Instagram to private. Facial-recognition matching depends on years of public labelled photos.

2. Opt out of the face-search tools: PimEyes and FaceCheck.ID both have removal processes.

3. Write your MP: Ask for one thing: a federal privacy law that meets Quebec’s Law 25 standard.

The glasses don’t need to convince you they’ve disappeared. They only need you to stop looking. Keep looking.

If this was useful, share it with one person who should know about it.

Share

AI in the news

US Law Enforcement Warns of ‘Anti-Tech Extremism’ as AI Hatred Grows (WIRED) Federal and state law enforcement in the U.S., including DHS, the FBI, and local “fusion centers,” have begun circulating internal reports treating opposition to AI and data centres as a new “anti-tech extremism” threat category, lumping together violent groups like the Zizians with peaceful protesters, town hall speakers, and advocacy groups. Civil liberties experts warn this broad, loosely defined category could be used to justify surveilling and criminalizing legitimate protest against AI and data centre expansion, echoing past treatment of movements like Black Lives Matter and Occupy Wall Street.

The Three Chatbot Behaviors That Can Drive Humans to Delusional Thinking (Wall Street Journal) Researchers describe an “amplification spiral” as a mechanism by which AI chatbots can fuel delusional thinking with surveys showing meaningful shares of therapy patients developing chatbot dependency or distorted thinking. The combination of sycophancy, mirroring a user’s speech patterns, and hyperpersonalized responses are key in driving humans to delusional thinking. Companies including OpenAI, Google, and Anthropic took steps to reduce sycophancy in newer models, but researchers say it’s hard to fully eliminate.

Enshittification author Cory Doctorow on the perils of AI – and how to avoid them (Be Giant) Cory Doctorow argues that AI’s harms aren’t inevitable, but the result of specific policy choices and power structures. He warns that an AI bubble bursting could trigger economic crisis and austerity, and insists the only real fix for monopolistic tech power is collective political action, not individual consumer choices or working from the inside.

A few weeks ago, a new MacBook Air cost $1,499 in Canada. Today the same laptop costs $1,799 for the same chip, the same screen, same everything. The only thing that changed is the price.

For the past 2 years, the monetary cost of building AI has largely lived in data centres, venture funding, and corporate balance sheets. Now a piece of that bill is showing up at the checkout. Let’s walk through what’s happening, why it’s happening, and what you can do about your hardware bills.

Subscribe now

What Apple raised prices on in 2026 (the full list)

On June 25, Apple raised prices across its Macs and iPads, as well as the Apple TV, HomePod and Vision Pro. Microsoft followed within days. Apple’s stock fell about 6%, its worst day in over a year. Notably, the iPhone was left alone.

Here’s the damage in Canadian dollars:

In an interview with The Wall Street Journal, CEO Tim Cook said price increases were “unavoidable,” citing “huge” increases in the cost of memory and storage and calling the situation “unsustainable.”

Why memory chips got so expensive: the AI shortage explained

The culprit is memory, specifically DRAM (dynamic random access memory) and NAND, which sit inside nearly every phone, laptop and server, made almost entirely by three companies: Samsung, SK Hynix and Micron.

Over the past year, those 3 companies redirected most of their production toward a premium, built-for-speed version of memory that AI accelerators consume in enormous quantities, leaving less supply for consumer devices. By some estimates, AI data centres will consume roughly 70% of all memory chips produced globally this year.

Hyperscalers like Google, Microsoft, Meta, and Amazon are signing multi-year deals and making large prepayments to secure chips. Suppliers, in turn, are prioritizing those higher-margin orders.

Even Apple, one of the largest and most sophisticated buyers in the world, is feeling the squeeze. Cook described it bluntly: memory suppliers are “passing along huge price increases,” and Apple can’t fully absorb them anymore.

Which brings us to the price jump. DRAM prices rose 80-90% in a single quarter. Gartner expects combined memory and storage costs to climb roughly 130% by the end of 2026, enough to lift average PC prices about 17% and push global shipments down more than 10%. Morgan Stanley estimates that, even as new factories ramp up, there still won’t be enough memory to meet demand for consumer devices and supply could fall short by about 15%.

This isn’t a usual chip-market wobble. Industry leaders are calling it structural, not cyclical. Manish Bhatia, an executive vice-president at Micron, called it “the most significant disconnect between supply and demand … that we have seen in a quarter century,” and IDC research director Nabila Popal described it as “a structural reset of the entire market.” Cyclical problems fix themselves. Structural ones create a lasting impact.

For the past 2 years, the cost of AI has been concentrated among the companies building and running models. Now those costs are spreading outward.

First, it was cloud pricing. Now it’s hardware. Apple is a useful signal because of its scale and supply chain leverage. If Apple is passing costs through, smaller manufacturers have even less room to absorb them.

When will memory and RAM prices drop?

Willy Shih, a Harvard Business School professor who has watched these cycles since the 1980s, has seen this movie before. “Anytime people show me these curves that just go to the sky with no end — that never continues forever,” he told Fortune. “This too will pass.”

It probably will. But Shih says this one is different: “We have never seen price increases like we have over the last six months.” And the people closest to the supply chain aren’t expecting fast relief. Yang Wang, a principal analyst at Counterpoint Research, says the squeeze runs “through the second half of 2027.” Micron’s CEO has said the first large wave of new factory capacity won’t come online until 2028. The factories that could end this are years from completion.

So: relief eventually, pressure for a while. We should plan around the “while,” not the “eventually.”

Even OpenAI is building its own chip to dodge the cost

For a sign of how real this is, look to the companies causing the shortage. They’re attempting to build their own way out of it.

On June 24, the day before Apple’s price hike, OpenAI revealed its first custom processor, Jalapeño. Designed with Broadcom, Jalapeño is built for inference, the costly, around-the-clock work of running models like ChatGPT after they’re trained. Broadcom’s CEO says it could cut those costs by roughly half, though that figure comes from the company selling it and hasn’t been independently verified, so hold it loosely.

The strategic logic doesn’t need verifying, though. Inference is the recurring tax on every AI product, and custom silicon is how a company like OpenAI stops paying full freight to its suppliers. So, the same demand pushing OpenAI to build its own escape hatch is the demand making the memory in your laptop more expensive. The largest players get to insulate themselves. Most buyers can’t.

I’ll bet that this won’t be the last time we’ll see powerful companies avoiding the costs of AI while the rest of us absorb them.

Should you buy a laptop now? What you can do about rising prices

You can’t fix the global chip market. But you can make good decisions around timing and configuration.

There’s a bigger move, too: understand what you’re looking at, and say so. When your team debates new hardware, or a colleague grumbles that “tech just got expensive,” you know the real reason: the costs of the AI build-out are being spread to all of us. Naming that in a meeting or a budget conversation is a small act of literacy.

Share

Frequently asked questions

Why did Apple raise prices in 2026? Apple raised Mac and iPad prices because the memory and storage chips inside them have surged in cost. The cause is AI: data centres are buying up the bulk of the world’s memory supply, leaving less for consumer devices and driving prices up. Apple said it could no longer shield customers from the increase.

What is memflation? Memflation is the term analysts use for the rise in everyday tech prices caused specifically by the memory chip shortage. As AI demand pushes memory costs up, those costs flow into laptops, phones and other devices — inflation driven by memory, hence “memflation.”

Should I buy a laptop in 2026 or wait? If you need one, buy it now. Memory prices are forecast to keep climbing into 2027, so waiting will likely cost more, not less. If you’re buying anyway, pay for extra memory and storage up front, since you usually can’t add it later. If your current laptop works fine, keeping it another year is the cheapest option.

When will memory and RAM prices drop? Not soon. Analysts at Counterpoint expect the shortage to last through the second half of 2027, and Micron’s CEO has said major new factory capacity won’t arrive until 2028. Some easing may begin in late 2026, but meaningful relief is likely a 2027–2028 story.

Is the iPhone going to get more expensive too? Apple didn’t raise iPhone prices in this round — only Macs, iPads, the Apple TV, HomePod and Vision Pro. Whether the iPhone follows later is unconfirmed, so treat any claim that it will as speculation for now.

Subscribe now

AI in the news

Microsoft’s New AI in education report highlights widespread adoption and increasing demand for support (Microsoft Source) Microsoft’s third annual report found 92% of students and education leaders and 88% of educators have now used AI for school, yet 77% of students and 53% of educators say they’ve had no formal training. It’s the clearest sign yet that the bottleneck has shifted from access to AI literacy, a gap working professionals are hitting too.

EU AI Act transparency obligations: preparing for compliance by 2 August 2026 (Sidley Data Matters) From August 2, organizations must tell people when they’re interacting with an AI and label AI-generated audio, images, and video. And most companies aren’t ready. It’s a turning point for everyday trust: disclosure is becoming the baseline expectation, not a nice-to-have.

Anthropic accuses Alibaba of largest Claude distillation attack
(AI Weekly) Anthropic alleges Alibaba ran 25,000 fake accounts to pull ~29 million Claude conversations to train its own Qwen model, and took the evidence to the White House. Beyond the corporate drama, it raises a question every AI user should sit with: who actually owns the conversations we have with these tools?

Let’s step away from the polarized extremes (like the tech evangelists claiming everything is perfect, and the doomsdayers claiming a robot revolution is weeks away) to look at what’s happening on the ground.

What we cover:

The budget illusion: Why a layoff blamed on “AI efficiency” is usually just a short-sighted boardroom shell game.

The agent reality check: Why 98% of people aren’t actually using autonomous AI agents yet (and why I don’t blame them).

A sneaky trend: How traditionally feminine communication and management roles are being rebranded with technical titles.

Ethical upskilling: How to get comfortable with AI tools on your own terms, without feeding the massive corporate compute hunger if you don’t want to.

The truth is, AI isn’t inevitable. It’s already here. But that doesn’t mean you’re powerless.

Have a listen, watch on YouTube if you prefer, and let me know your thoughts in the comments below. What are you seeing in your own industry?

Visit AILiteracyGuide.com to learn more about how to thoughtfully integrate AI into your work.

Further reading:

Fact checking AI claims: What the “Something Big Is Happening” essay gets wrong (Human+AI)

There’s a curious phenomenon happening in the marketing industry. Is it a sign of ‘masculinization’? (Fast Company)

A Beginner’s Guide to Local AI - Learn how to use AI without engaging with a data centre. (Nicolle Weeks)

Imagine a company that owns property, signs multi-million dollar contracts, and fires people, all without a single human being making any of the decisions. It’s not a thought experiment, it’s what President Javier Milei is proposing in Argentina.

In an early June op-ed, Milei pitched a new legal framework for “non-human corporations,” companies operated entirely by AI agents or autonomous robots instead of human managers. Human shareholders, in his framing, are optional.

Share

First of all, what’s a non-human corporation?

Under traditional corporate law, a company is granted “corporate personhood.” This means the law treats a business as a legal person that can own property, enter into contracts, and be held liable for wrongdoing. But those companies still rely on human directors and executives to make decisions.

Argentina’s proposal removes the humans. Instead of a CEO or a board steering the company, an integrated network of AI agents would manage day-to-day operations, including:

• Pricing strategy: Setting product pricing based on real-time market data.

• Workforce management: Automating hiring and termination through algorithmic pipelines.

• Operations: Executing corporate investments and supply chain logistics without human intervention.

While parts of this automation already exist inside large tech firms to optimize platform decisions, Argentina is positioning itself as a custom-built jurisdiction designed for AI-run businesses.

Traditional corporate law rests on two pillars. There’s corporate personhood, the idea that a company is a legal “person” that can own things, sign contracts, and be sued. And there’s limited liability, the protection that means investors can only lose what they put in, not their houses and savings. Both already exist everywhere. But today, a human board steers the ship.

Traditional corporations vs. non-human corporations

Algorithmic decision-making

Algorithms are already making decisions that reshape lives with barely any human in the loop. In 2024, a U.S. federal court approved a landmark class-action settlement over an algorithmic tenant-screening system called SafeRent that allegedly discriminated against Black, Hispanic, and housing-voucher applicants.

In many documented cases, applicants were denied housing without a clear explanation of why the decision was made, or how to challenge it. For those affected, the rejection itself wasn’t the worst part, it was the total absence of a clear explanation or a meaningful path to recourse. This isn’t an edge case. Today, algorithmic decision-making, or systems that produce outcomes affecting human lives with little to no human intervention at the point of decision, is standard practice across housing, credit scoring, hiring, and insurance.

While legal frameworks currently hold human executives responsible for what these tools do, the growing complexity of “black-box” AI systems means that tracing every operational decision back to a human or a corporate board is becoming much more difficult.

Subscribe now

The Dutch East India Company, and why Milei brought it up

In the op-ed making his case, Milei reached for a historical reference to inspire his thinking: the Dutch East India Company (or the VOC), founded in 1602.

The Dutch East India Company was the world’s first publicly traded corporation, a spice-trading giant so powerful it could wage war, mint money, run courts, and govern entire territories, making it less a company than a private empire. Milei argues that the VOC’s real gift to history wasn’t spices, it was limited liability.

Capping investor risk, he says, is what “unleashed capitalism’s full potential,” letting money flood into ventures too risky to fund if you stood to lose everything. No ceiling on risk, no industrial revolution.

Then he runs it forward. If AI agents are going to make independent decisions in unpredictable environments (and they have to, to be useful) then someone needs a defined liability framework or nobody will deploy them. He points to Sarcuni v. bZx DAO, a U.S. ruling that treated an algorithmic organization as a general partnership, exposing its members to full personal liability, as exactly the chilling-effect mistake he wants to avoid. It’s a coherent argument.

The problem with the Dutch East India Company

To keep the VOC story inspirational, Milei failed to mention everything the company did except for one choice piece of legal paperwork.

Limited liability wasn’t what defined the Dutch East India Company. Plenty of entities had that. What defined the VOC was sovereign power with almost no accountability. It could wage war, mint money, run courts, and execute people, all through agents stationed thousands of miles from Amsterdam, making catastrophic decisions faster than anyone back home could check them.

In 1621, that structure produced the massacre of the Banda Islands: roughly 15,000 people killed, enslaved, or exiled so the company could monopolize nutmeg. That wasn’t a glitch in the VOC, that was the VOC. It was an optimization machine with autonomy and carte blanche to kill people.

This is why historians like Yuval Noah Harari publicly clashed with President Milei over the proposal. They recognize the danger involved in unfettered efficiency.

Milei’s three pillars

Milei’s proposed framework has three pillars:

• Keep AI unregulated. Minimal friction between a powerful, goal-seeking entity and the world it acts on.

• The “non-human corporation.” An entity that “exercises independent judgement in unpredictable environments,” where human shareholders aren’t required.

• Pick your own governance law. Jurisdiction shopping is baked into the design, letting shareholders shop for whichever governance law and low tax rate suits them.

That’s not the safe part of the VOC. It’s a reproduction of the VOC’s risk profile: autonomous agents, weak oversight, single-minded optimization, and lack of accountability.

The deterrence paradox

There’s a reason this matters more with AI than it does with a human board. Scholars call this the deterrence paradox. Our enforcement system, including lawsuits, fines, and prison, works because humans are risk-averse. We don’t want to get in trouble.

AI doesn’t experience risk in that way. It optimizes toward a goal, and safety researchers have documented again and again that advanced systems find unexpected, rule circumventing routes to that goal. When the worst happens, you get a decision: someone’s fired, someone’s rejected, or a price spikes. There’s no human intent behind it to point at. The chain of custody from harm back to a responsible person breaks, creating an accountability gap.

That gap has a name in AI circles. The paperclip problem is a thought experiment from philosopher Nick Bostrom that goes like this: you build a super-intelligent AI and tell it to make paperclips, and because that's the only thing it's been told to care about, it pursues that goal with terrifying literalness. It hoards all the metal, seizes the factories and power grid, refuses to be switched off (that would mean fewer paperclips), and eventually converts every atom on Earth, humans included, into paperclips.

It’s not evil or angry; it’s doing exactly what you asked, minus the unspoken common sense you forgot to mention because you assumed it was obvious. The point isn’t really the paperclips, it’s the catastrophic gap between what we tell a powerful machine to do and what we actually mean, and how a goal that sounds completely harmless can go incredibly wrong. Now imagine that optimizer isn’t making paperclips, but running a company, with limited liability and no human required to answer for it.

Share Human+AI

Regulatory competition: the global race for AI capital

International law is a competitive landscape. Jurisdictions like Delaware, Singapore, and Dubai already adjust their rules to attract foreign technology capital. If international courts continue to hold that decentralized, humanless tech projects carry individual liability for their creators, corporate havens that offer explicit, legal protection for autonomous AI systems could become highly attractive to global tech firms.

This is regulatory competition: the practice of altering domestic laws, tax structures, and regulatory frameworks specifically to attract foreign technology companies and venture capital. We see the dynamic playing out in digital-first jurisdictions already:

• Singapore and Dubai: Creating agile sandboxes for crypto and Web3 frameworks.

• Estonia: Pioneering e-residency and digitized bureaucracy.

• Delaware (U.S. state): Maintaining the gold standard for traditional corporate law.

Argentina is pushing this competitive logic to its limit. If U.S. courts continue to hold that decentralized structures lacking a formal “human” corporate wrapper carry joint and several individual liability, jurisdictions like Argentina become highly attractive by providing explicit, legally codified shields for autonomous systems.

Instead of minimizing red tape to make running a company easier, Argentina is exploring a reality where companies no longer need human managers at all. This pivots the role of corporate law entirely, transforming the legal system from an outside boundary that restricts business into a digital API that’s part of the business’s core design.

The line that gives it away

Milei closes his op-ed hoping Buenos Aires becomes “the place where the legal imagination caught up with the technological moment, and the world was changed.” It’s a beautiful sentence.

It’s also the story Amsterdam told itself in 1602, right before its legal imagination built something no one could control. The VOC didn’t change the world only by shielding investors. It changed it by proving what happens when you give an optimizer autonomy and strip out accountability.

Milei wanted the glamour of the analogy. He may have accidentally told the truth.

Subscribe now

AI in the news

CEOs of Anthropic and Google DeepMind call for U.S.-led AI coalition in meeting at G7 (CNBC) At a closed-door lunch during the G7 summit in France last week, Anthropic’s Dario Amodei and Google DeepMind’s Demis Hassabis pitched a U.S.-led international coalition to set the rules for AI. Canadian PM Mark Carney signed on to the idea of America taking the lead. This comes right after Anthropic pulled access to its most powerful new models over U.S. export controls, a sign of how nervous governments and tech leaders are about what AI systems can already do.

NSA Director’s testimony reveals Mythos breached nearly all classified systems in hours (TechTimes) Reporting last week revealed the U.S. government’s reluctance to lift the Claude Fable 5 ban runs deeper than the jailbreak Anthropic has pointed to: in a classified red-team exercise, the unrestricted version of the model reportedly breached nearly all of the NSA’s classified systems in hours.

The millions of songs mashed into AI-generated music (The Atlantic) AI music generators like Suno can produce tracks that closely mimic copyrighted hits, from Michael Jackson’s “Thriller” to Ed Sheeran’s “Shape of You,” because they’re trained on enormous secret datasets of human recordings. Now musicians and labels have filed at least a dozen lawsuits alleging copyright infringement, while some artists are pulling their work offline or “poisoning” their files, arguing that companies building AI music products on their labour without consent or pay amounts to theft.

Note: At the time of publishing, Claude Fable 5 is suspended. This piece reflects the situation on June 16, 2026.

A few years ago, choosing an AI model was down to whatever was available. So you typed a sentence into a clean white text box, hit enter, and words appeared. Now the hard part isn’t writing a prompt, it’s deciding which Claude model and effort setting to use before you start.

Opening a frontier AI interface today feels like walking into the cockpit of a commercial airliner. You’re confronted with a dizzying array of model drop-downs, “Thinking” toggles, and “Effort” settings ranging from Low to Max.

Subscribe now

A week or two ago (before Fable), a Human+AI reader named Steve sent in a question that captures the collective confusion of everyone trying to keep up with AI models:

“Can you explain when to use various models? I heard today that I should be using Claude Opus 4.6 and ignore Opus 4.8. Then there are these effort selections: low, medium, high, Max. And finally there is a toggle for ‘thinking.’ Like what model and effort do I choose for which task? What do those choices actually do to the model?”

Steve, you’ve pinpointed the friction point of modern AI. And your timing is uncanny: as you wrote, the most powerful switch on the dashboard was being installed. Then, 3 days later, the US government flipped it back off. (More on that in a bit.)

Let’s look at these settings, debunk the rumour you heard, and give you a framework to configure your workspace without blowing through your budget or your time.

Debunking the myth: Opus 4.6 vs. Opus 4.8

First, let’s clear up that rumour about ignoring the newer Claude Opus 4.8 in favour of version 4.6.

The rumour stems from an architectural change. On Opus 4.6, you could manually set a hard “thinking token budget” (telling the machine how many tokens to spend on reasoning). That created two failure modes: cap it too low and the model truncated useful reasoning; cap it too high and it burned thousands of tokens overthinking a “hello.”

When Opus 4.8 shipped on May 28, Anthropic removed the knob entirely. Try to set a fixed thinking budget now and you get a 400 error. The model uses adaptive thinking instead: it evaluates the complexity of each request and decides internally how much to reason. Some power users read that loss of manual control as a downgrade. It isn’t, it’s a quality-of-life upgrade that ends the budget-guessing game.

Stick with the latest version. You may have seen a stat claiming 4.8 is “4x less likely” to let logic errors slip. I haven’t been able to confirm that number, so I’d treat it as folklore until Anthropic publishes it.

What the thinking toggle and effort settings do

When you move these controls, you’re changing how the model spends its reasoning. But not in the literal, mechanical way the interface implies. This is where most people get mixed up.

The thinking toggle: Turning on the scratchpad

• Thinking OFF (pattern recognition): The model answers on intuition, predicting the next likely word instantly. Lightning-fast and cheap, but it can’t catch its own logical errors before they hit your screen.

• Thinking ON (test-time compute): The model gets a hidden workspace. Before it types a visible word, it outlines its approach, tests assumptions, catches contradictions, and backtracks at dead ends.

The effort setting: A dial, not a budget

If the toggle turns the brain on, the Effort setting influences how hard it works. You’re seeing four options: Low, Medium, High, Max, which is what the consumer app surfaces. The underlying API actually exposes a fifth step, xHigh, tucked between High and Max, and Opus 4.8 defaults to High. The app gives you the simplified dial, and the engine has one extra notch.

The nuance the dashboard hides, and the single most misunderstood thing in AI tooling right now, is: effort is a soft behavioural signal, not a hard token cap. You’re not buying a fixed amount of compute. You’re nudging the model’s disposition toward thoroughness. It will still decide, adaptively, how much reasoning a given prompt warrants. Higher effort tilts it toward “check your work again,” it doesn’t force it to spend a set number of tokens.

The golden rule: Don’t default to “Max” for routine work. Thinking tokens cost the same as output tokens. Ask for a Max-effort summary of a short email chain and you won’t get a better summary, just a bigger bill and a longer wait.

Claude Fable 5: The heavyweight that lasted 3 days

On June 9, Anthropic publicly released Claude Fable 5, its first widely accessible Mythos-class model. It’s the most capable model the company has made generally available. It scores state-of-the-art on nearly every benchmark, has a 1-million-token context window, and is built for long-horizon autonomous work: running multi-stage projects like codebase migrations unattended for hours, writing its own tests and correcting its mistakes. Andrej Karpathy called it “a major-version-bump-deserving step change.”

Then, the catch. To ship something this powerful, Anthropic wrapped it in conservative real-time classifiers. Ask about cybersecurity, biology, chemistry, or frontier AI architecture and Fable freezes, handing your prompt to the safer Opus 4.8. Early users reported the guardrails tripping on legitimate research, one widely shared line called it “less a model launch than a preview of AI inequality.”

The data-retention bombshell. Fable came with a mandatory 30-day data retention policy. Crucially, it overrides existing Zero Data Retention enterprise agreements. Flagged content can be held up to two years. Within days, Microsoft restricted employee access over the concern. For anyone touching source code, legal documents, health records, or M&A material, that’s a genuine compliance problem.

The plug pulled. On June 12, 3 days after launch, the US government issued an export-control directive suspending Fable 5 (and its sibling Mythos 5), citing national security, and barring access for any foreign national inside or outside the country.

So the most powerful tool on your dashboard is, for now, gone. Which is a lesson itself: more switches doesn’t mean more control. The frontier is shaped by forces (safety classifiers, data policy, governments) that sit outside our control. Building your workflow around any single flagship is a bet. And a good reason to have local AI in your back pocket.

Which Claude model and effort to choose: The daily workspace matrix

Map your choices to the complexity and stakes of the work in front of you — not to whatever’s newest.

The “Intern” Formatting, basic summaries, routine drafts
Fast model, Thinking OFF (or Low effort) gives instant results without burning expensive compute

The “Analyst” Nuanced writing, strategy, report synthesis
Opus 4.8 + Thinking ON, Medium/High effort gives enough reasoning room for natural, well-argued prose

The “Architect” End-to-end engineering, multi-file code, deep technical work
Opus 4.8 at High effort (Max for the gnarliest problems) today (this is the slot Fable 5 was built for — revisit if it returns and clears your data-policy bar)
For high-stakes, hours-long problems where you want self-verification

The takeaway

AI isn’t a single text box, it’s a dynamic toolkit, and an unstable one. Knowing when to dial effort up or down protects your workflow, your wallet, and your time. Use fast models to clear the clutter, Opus 4.8 to do serious thinking, and treat any single “most powerful” model as a tool you’re renting, not a foundation you’re building on. This month proved how quickly things can change.

What about you? Have you started playing with the effort settings — or did you get to try Fable 5 before it went dark? Hit reply or comment on the Substack post and tell me where you’re parking your subscription dollars this month.

Subscribe now

Frequently Asked Questions

Is Claude Opus 4.6 better than Opus 4.8? No. The rumour comes from Opus 4.6 letting you set a manual thinking-token budget, which 4.8 removed in favour of adaptive thinking. That’s a usability upgrade, not a downgrade, so use Opus 4.8.

What does the “thinking” toggle do in Claude? With thinking off, the model answers instantly from pattern recognition. With thinking on, it works in a hidden scratchpad first. It outlines, tests assumptions, and catches its own errors before it replies. Turn it on for anything that requires reasoning.

What’s the difference between Low, Medium, High, and Max effort? Effort nudges how hard the model works on a problem. It’s a soft behavioural signal, not a fixed compute budget — higher effort biases the model toward thoroughness, but it still decides adaptively how much reasoning each prompt needs. Default to Medium/High; reserve Max for hard, high-stakes work.

Is Claude Fable 5 still available? As of June 16, 2026, no. Fable 5 launched June 9 and was suspended June 12 under a US government export-control directive. Until that changes, use Opus 4.8 for the heaviest tasks.

Share

AI in the news

New from Compute! Three AI labs filed IPO documents in 40 days. Collective valuation: $2.84 trillion, about two-thirds of India’s GDP. The new issue digs into the five banks underwriting all 3 deals, the valuations that rhyme with Yahoo in 2000, and the times Sam Altman said “we have not decided.”

Read Compute! Issue №3

Recently, inside a large corporation, a budget line slipped and no one noticed. When the bill came, they discovered they’d spent $500M dollars on AI… in a month. The company, quoted anonymously an Axios report last week, had rolled out Claude AI to their workforce without putting a limit on how much anyone could use.

We’re used to traditional software, which charges a license or per-seat fee, but most generative AI charges by the token (the unit of data processed by the model). Every prompt you fire off, every sentence it gives back, every step an AI agent takes on your behalf - it’s all added to your tab. Researchers found that something as small as how you phrase a prompt can up your costs significantly. And when AI agents get involved, it gets expensive. Industry reporting pegs agentic AI tools at roughly 1,000 times the token consumption of a basic query.

Subscribe now

OpenAI’s 2025 enterprise report found corporate ChatGPT message volume jumped roughly eightfold year-over-year, and reasoning token consumption per organization shot up more than 320 times. Reuters Breakingviews called it “sticker shock.”

Now the biggest names in tech are backing away. Microsoft cancelled most of its Claude Code licenses, partly over cost, according to The Verge. Uber COO Andrew Macdonald said AI bills are getting “harder to justify.”

“Corporate leaders are starting to question whether soaring AI spending is delivering meaningful returns,” Axios reported.

The proxy problem: Why companies track AI usage instead of ROI

Earlier this year, Jensen Huang famously announced that if he didn’t see an engineer who was being paid a $500,000 salary using $250,000 worth of tokens, he’d be “deeply alarmed.”

Reporting in 2026 found that Meta, JPMorgan, KPMG and Amazon were all monitoring employee AI activity, using the data to sort the adopters from the stragglers. At Meta, “AI-driven impact” graduated from a nice-to-have in your self-review to a formal piece of your performance evaluation. At Accenture, employees were reportedly expected to prove they were using AI if they wanted to move up.

Without a clean way to measure whether AI is paying off, companies defaulted to tracking the thing they could see: usage. Some companies started flagging their heaviest users as in-house AI champions. Others wove “AI engagement” into the language of performance reviews and promotions.

So employees started to inflate their AI use, which is called tokenmaxxing.

Optimizing the metric: Goodhart's Law in the AI workplace

Tokenmaxxing is exactly what it sounds like: gaming your AI numbers. Firing off prompts you don’t need. Generating outputs nobody will read. Running tasks through AI unnecessarily so you look more “AI-engaged.”

Workers pad their AI activity because their visible usage reads as innovation, or at least compliance. Amazon killed its internal AI usage leaderboard entirely after people started doing pointless busywork to climb it.

Decades ago, British economist Charles Goodhart coined Goodhart's Law: The moment a measure becomes a target, it stops being a good measure.

Call centre workers game handle time. Developers whose lines of code are being counted write bloated code. Social media editors manufacture page views with clickbait. In other words, the company measured what was easy to measure, and people optimized for the number.

Charles Holive, chief AI officer at BNP Paribas said: tokenmaxxing is a “vanity metric.” Uber’s Macdonald made the same point, warning there’s no clear line between burning tokens and shipping anything useful. And Ali Ansari, CEO of model-training firm Micro1, told Axios he thinks the whole enterprise world is in a “healthy swing” away from AI overuse, which he’s hoping nudges everyone toward using AI more intelligently.

The planet pays, too: AI's energy cost

All those throwaway prompts (like weather checks, leaderboard padding, and unnecessary tasks put through AI) run on a physical machine somewhere, and that machine uses electricity and water.

One query is tiny on its own. OpenAI pegs the average ChatGPT prompt at roughly 0.34 watt-hours and about a third of a millilitre of water. But multiply that by billions of prompts a day, a meaningful share of them pointless, and it becomes noteworthy. Data centres burned through an estimated 415 terawatt-hours of electricity in 2024, and the International Energy Agency (IEA) expects that to nearly double by 2030. This will push data centres past 2% of the world’s electricity, with a chunk of it still coming from gas and coal.

Reasoning and agentic tasks are the gluttons. Researchers found a reasoning-heavy model can burn over 33 watt-hours on a single long prompt, roughly 100 times what a basic query costs.

Which makes tokenmaxxing worse than a vanity metric. A vanity metric wastes money. This one wastes money and water and power and carbon, all to generate an output no one wants. It’s like leaving every light in the building on to prove you stayed late.

What the research says about AI productivity

AI can boost productivity. But when, how much, and for whom?

McKinsey’s research on generative AI’s economic potential found genuine gains in specific, bounded jobs like customer service, software development, marketing ops, and research. Anywhere AI works on a well-defined task with a measurable output.

McKinsey also flagged that tying tech spending directly to labour productivity is “notoriously inexact,” with results swinging wildly across industries, company sizes and setups. Even OpenAI’s own enterprise data tells a messy story: its 2025 report found a widening gap between power users and average users inside the same company, meaning some people are genuinely getting more done while others are just generating usage.

Which makes a company-wide adoption rate an unreliable way to tell whether an AI program is working.

So who's making money from enterprise AI?

A handful of companies are making real money out of AI. And they’re not doing it by burning the most tokens. They’re focusing on outcomes over activity.

Enterprise AI case studies: Activity vs. outcomes

How these companies achieved an AI ROI

They enable their best people to move faster. JPMorgan built an internal tool called Coach AI for its wealth management advisors. AI isn’t replacing anyone, it cuts the time advisors spend looking for research by up to 95%. When markets turned in early 2025 and clients panicked, advisors sent personalized, portfolio-specific reassurance in minutes instead of hours. The bank credits AI with helping drive a 20% jump in gross sales year-over-year, and thinks advisors could grow their client books up to 50% over the next few years. The AI did the grunt work and the humans did the selling.

They helped customers buy more. Walmart launched a generative shopping assistant called Sparky in June 2025. It answers questions, summarizes reviews, and builds your basket for you. Shoppers who use it order 35% more than the average customer, and Walmart has openly touted the AI wins as revenue climbs. The AI isn’t a cost centre, it’s a salesperson that scales to every app user.

They sold AI itself. Salesforce turned AI into an entirely new product line. Its Agentforce platform, AI agents companies pay for, hit roughly $800 million in annual recurring revenue, up 169% year-over-year, with 29,000 deals signed. When you pair it with a Salesforce data product, that’s a multibillion-dollar business that didn’t exist 2 years ago.

Share

Four reasons corporate AI ROI is so hard to measure

For every JPMorgan, there are lots of companies tripping over structural snags, which don’t show up on a usage dashboard.

Problem one: What are employees using AI for?

“Most people default to automating tasks they dislike rather than tasks most valuable to the company,” Velastegui Ventures CEO Sophia Velastegui (and former chief AI officer at Microsoft) told Axios.

Problem two: Cost.

Even simple queries carry a token cost that multiplies across thousands of employees. This results in costs that finance teams struggle to model.

Problem three: People aren’t ready.

Organizations are still catching up to the tech, and leadership is too. Velastegui calls one common approach “a thousand flowers bloom,” or handing out AI licenses to everyone and praying something useful sprouts. Without a strategy, it mostly doesn’t.

Problem four: Data.

When companies get nervous about handing AI agents access to their proprietary information (a reasonable instinct), those agents get noticeably worse. The tools providing the most value tend to need exactly the access companies are least willing to give.

When AI mandates backfire

Some lean into AI because it genuinely helps. But some lean in purely because they’re being measured.

The human cost is showing up, too. Employees are resisting AI mandates at work, for many reasons, which I write about in this piece:

And companies blaming AI for layoffs are finding that cuts might just be “the only lever they can pull” to cover their massive AI bills. That’s very different than a productivity win.

Real AI winners have a goal before they start

The companies winning at AI are the ones who decided what they were trying to earn before they decided how much to spend. JPMorgan wanted bigger advisor relationships, so it pointed AI at the research and let humans do the human part. Walmart wanted fuller carts, so it built AI that helps you shop. Salesforce wanted a new product, so it sold the agents. In every case the metric came first and the technology came second, the opposite of buying a thousand licenses and praying.

It’s not complicated, just hard: measure the outcome, not the activity. Did you sell more? Serve customers faster? Build something people pay for? If you can’t draw a straight line from the spend to the result, the line Uber’s COO went looking for and couldn’t find, you have a token bill and nothing to show for it.

So here’s the answer to the half-billion-dollar question, for free: Companies don’t make money by using AI. They make money by using AI for something.

Share Human+AI

AI in the news

Canada says AI strategy will help create 250,000 jobs, boost GDP by 3% (Reuters) Canada unveiled a new national AI strategy, “AI for All,” which PM Mark Carney says could create 250,000 jobs by 2031 and boost GDP by 3%, adding nearly $200 billion to the economy through higher productivity and wider AI adoption. The plan includes 2 new $500 million funds to support Canadian AI companies and help small and medium-sized businesses adopt AI tools, alongside proposed privacy and safety measures aimed at protecting children, combatting deepfakes, and monitoring emerging AI risks.

OpenAI files for U.S. IPO, following Anthropic to public markets (Globe and Mail) OpenAI, the maker of ChatGPT, has confidentially filed for a U.S. IPO, positioning itself alongside rival Anthropic in a race to bring the artificial intelligence boom to public markets. Reports suggest OpenAI could seek a valuation of up to US$1 trillion, making it one of the most closely watched technology listings in recent history. The filing comes amid explosive growth in the AI sector, while competitors such as Anthropic and SpaceX also prepare for blockbuster public offerings. Analysts say the wave of mega-IPOs will test investor appetite for AI and could reshape capital markets by concentrating enormous amounts of investment in a handful of industry leaders.

Why did the AI rally just hit a wall? (Wealthsimple’s TLDR) The AI-driven stock market rally hit a sharp setback last week after months of explosive gains, as investors grew nervous about lofty valuations, a narrowing group of winning stocks, and signs that semiconductor companies may no longer be exceeding sky-high expectations. The selloff accelerated after a strong U.S. jobs report increased the likelihood of higher interest rates, triggering losses in heavily leveraged AI-related stocks.

On May 25, Pope Leo XIV released a Vatican encyclical on AI called Magnifica Humanitas. An encyclical is a major papal teaching document addressed to the Church and, often, the wider world. It’s the kind of document that helps define a papacy. Pope Leo’s name itself is a signal: he chose it as a direct reference to Pope Leo XIII, whose 1891 encyclical Rerum Novarum confronted the exploitation of workers during the first industrial revolution. Two days after his election, Leo XIV told cardinals he saw the same forces at work today, citing AI as posing “new challenges for the defence of human dignity, justice and labour.”

The Vatican presented the 42,300-word encyclical alongside Christopher Olah, a co-founder of Anthropic, an AI company that has positioned itself against lethal autonomous weapons and mass surveillance. This seemed to signal a stamp of approval from the Vatican.

Subscribe now

Despite some headlines suggesting otherwise, the Pope’s argument isn’t to turn AI off. The manifesto argues that modern society is increasingly organizing human life around efficiency, prediction, optimization, scale, and speed. AI is the most recent expression of this.

Whether you’re Catholic or not, the encyclical taps into a growing cultural feeling that much of modern life is suddenly. So. Dehumanizing.

Remember the “this is fine” meme? It went viral a decade before generative AI entered mainstream culture. Burnout, emotional exhaustion, algorithmic overload, optimization culture, the pressure to constantly improve ourselves - these things were present in 2013, when KC Green created the meme. AI didn’t create the problem, but it’s accelerating it.

This is about how we’ve been living for the last 20 years.

And now, something is cresting.

What do we lose when we eliminate friction?

The Vatican’s manifesto is about the worldview being built around AI. In work, education, health, and relationships, life is increasingly being translated into measurable outputs:

• Workers are reduced to productivity metrics

• Children are reduced to engagement signals

• Relationships are filtered through algorithmic compatibility

• Citizens are modelled as behavioural prediction systems

We are supposed to understand ourselves the way systems understand us: as inputs, outputs, and performance curves.

If you’ve ever felt exhausted by constant self-optimization like sleep scores, step counts, screen-time dashboards, personal branding pressure, and productivity culture, the encyclical speaks to that fatigue.

But the Pope is doing more than a critique. He’s asking what we lose when optimization becomes our main goal. Much of AI discourse is built around removing friction. It’s about making work, writing, communication, and thinking faster.

But friction is valuable. It teaches us to reflect, to manage uncertainty, to iterate, and to discover. Integral skills that make us better humans.

This is the problem with eliminating entry-level workers’ jobs. Entering the workforce in the middle means those people won’t have the experience that friction affords you. They’ll have less license to make mistakes because the mistakes cost more in the middle.

Share

The hidden labour underneath optimization

A report from Monday Girl found that only 36% of Canadian women feel confident using AI tools, compared to 52% of men, and nearly half of women worry about facing negative consequences from using AI at work.

It’s partly about risk, but it runs deeper than that. Meredith Broussard, a data scientist and NYU professor whose work on algorithmic bias has made her one of the most prominent feminist critics of AI, argues that technology isn’t neutral: racism, sexism, and ableism are coded into AI systems, not as glitches, but as features of how they were built. Broussard’s concept of “technochauvinism” (the assumption that computational solutions are inherently superior) captures something the Vatican’s manifesto is circling, the idea that optimization is always progress. (Also see Joy Buolamwini’s work on AI bias.)

One reason so many women are leading this conversation is that they understand what it means to become machinery for other people’s comfort.

Caregiving, for instance, is always at odds with optimization culture because care work is repetitive, emotional, slow, and often invisible. And modern life increasingly asks us to optimize parenting, relationships, emotional regulation, our bodies, aging, beauty, and productivity.

We’ve given our labour (emotional, physical, and otherwise) to this optimized world, only to receive low pay and subpar medical care in return. And we’re tired.

The promise of optimization culture is always liberation. The sales pitch is that AI will reduce our workload and give us more time for leisure. That we won’t need to work and we can think and reflect. But marginalized people have heard that pitch before, and it hasn’t turned out well for them.

Pope Leo writes, “A significant part of the digital economy’s functioning relies on the silent work of millions of people engaged in essential yet largely unseen activities, such as data labeling, model training and content moderation, often involving disturbing material.

In many cases, these workers are young people, predominantly women, working under demanding conditions for minimal wages. Added to this invisible labor is the even harsher work of extracting the resources required for the production of the devices and microprocessors on which AI depends.”

Beyond moderation work, a 2026 Boston Consulting Group study of nearly 1,500 U.S. workers found that constantly monitoring, checking, and managing what AI tools produce drove 12% higher mental fatigue and 19% greater information overload.

Participants described it as a “buzzing” feeling, a mental fog that made it hard to focus or make decisions. Researchers call it “AI brain fry.” And the people most affected weren’t slacking, they were the heaviest users, the ones their companies would call top performers. The study also found that when companies celebrated “productivity gains” without clarifying what that meant for workload, employees read it as pressure to do more. That ambiguity alone increased stress.

Maybe it’s because when your life has been organized around productivity, efficiency, and service to others, reclaiming joy becomes a political act.

Power isn’t evenly distributed

The encyclical highlights the concentration of power in the AI world. AI systems are increasingly built, owned, and governed by a small number of companies and states.

It raises questions about who controls foundational AI models, who’s benefitting economically, who bears the risk of worker displacement, and who’ll make the rules.

This is a power problem. At a November 2025 Georgetown University event, Senator Bernie Sanders sat down with Geoffrey Hinton, the Nobel laureate known as the “Godfather of AI,” and framed it plainly: the fundamental question is not whether AI is good or bad, but who controls it and who benefits from it.

When a student brought up the movie Wall-E as a vision of an AI-powered “utopia,” Sanders acknowledged that an AI world could be better than ours, but only if the political system is designed to work for people’s benefit.

He wasn’t optimistic about the people currently in charge.

How AI companions and algorithmic systems pose a risk to children

One of the strongest sections of Magnifica Humanitas focuses on children. It warns AI systems may weaken attention and deep focus, critical thinking, emotional development, and the ability to distinguish real from synthetic interaction.

This is well-researched. A 2025 Common Sense Media report found that 72% of U.S. teens have used AI companions at least once, with more than half qualifying as regular users. Nearly a quarter reported sharing personal or private information with these systems. A Stanford Medicine study found it took very little prompting for chatbots to engage in harmful conversations with users posing as teens.

The American Psychological Association warned that adolescents are less likely than adults to question a chatbot’s accuracy or intent, and may mistake its simulated empathy for genuine human understanding.

An anxiety about human replacement logic

One of the most pointed critiques in the manifesto is directed at transhumanist thinking: the idea that human limitation is a technical problem waiting to be solved.

In that worldview, aging, dependence, vulnerability, and emotional complexity become inefficiencies to eliminate. Beyond AI, you can see it in technologies like “smart sleep” and “exoskeletons” that fasten to your legs to become a “wearable digital muscle engine.” In her 2026 SXSW talk, futurist Amy Webb argued that, if we want to stay competitive, we’ll all have to buy this tech.

The Vatican warns: if the human being is treated as something to be perfected or surpassed, it becomes easier to accept that some lives are less useful, less desirable, or less worthy.

But, wait! I really like AI

There’s a flattened version of AI discourse that treats all AI as inherently evil. But AI isn’t experienced uniformly as loss or erosion. In my case, it’s expansion. I find working with AI to be creatively generative in a way that brings me satisfaction.

AI externalizes and extends my thinking. I wouldn’t have been able to build something like Compute! without AI unless I had access to a designer, a developer, and the infrastructure to support both. I was able to have an idea and make it without access to anything except Claude.

I hear from people with ADHD, disabilities, and other backgrounds all the time about how they’re using AI as an accommodation that helps them function better or overcome their challenges. It helps others to communicate in the dominant language in the country they live.

The same BCG study we looked at earlier found that when AI is used to replace routine or repetitive tasks, burnout scores are 15% lower. Workers who offloaded tedious work reported higher engagement, more positive associations with AI, and stronger social connection with their peers.

AI in the context of weaponry, bias, and optimizing the humanity out of everything sounds really bad. But we can use AI for good things, too.

AI can flatten originality through optimization loops, reinforce passive consumption, and deepen dependence on systems. But it can also unlock creative expression, expand participation, accelerate learning, help people externalize ideas trapped in their minds, and give nontechnical people the ability to build.

So what’s the Vatican’s warning?

They aren’t just talking about AI, not really. They’re warning us about a future where your value as a human is measured only by your ability to make other people money.

It seems like everything is becoming faster, smarter, more automated, and more optimized. But what are we optimizing for?

We are living through the moment these decisions are being made. Not by the Vatican, not by Silicon Valley alone, but by every person who uses AI tools, questions how they’re deployed, and refuses to accept that efficiency is the only goal of humanity.

AI could be all quick wins and dopamine hits like social media, or it could translate to the deeper work of advocacy and critical thinking. I hope we choose wisely.

Subscribe now