Doctorow tells The Atlantic the AI bubble is set to burst

The frame Cory Doctorow keeps coming back to on The Atlantic's podcast is not which model beat which on which benchmark, but who is using whom. He splits AI users into two camps. A centaur, he says, is a person assisted by a machine. A reverse centaur is the inverse, a human conscripted to do the tasks the machine directs at its own pace. People who chose to adopt AI tend to like it; people who had it imposed on them find it useless or worse. That asymmetry is the lens he pushes readers toward in his new book, 'The Reverse Centaur's Guide to Life After AI: How to Think About Artificial Intelligence Before It's Too Late,' out in June 2026.

The economic argument behind the book is the one he has been building publicly for months and that he restates in plain numbers. The AI industry, by his accounting, is spending around $1.4 trillion against roughly $50 billion a year in revenue, and each additional customer loses the providers more money rather than less. He also told Democracy Now that when labor drives automation it is usually in service to making the product better, while when capital drives it the goal is usually to make more of the product, often by firing high-waged workers and replacing them with substandard algorithms.

The line he keeps returning to is that AI cannot do your job, but an AI salesman can convince your boss to fire you for an AI that cannot actually do the work. That is the political problem the book is really about. The question is not whether the models are getting more capable in the lab, it is who gets to decide whether the machine sets the pace or the worker does.

The honest caveat is that this is one well-argued case from a critic with a book to sell, not a settled forecast. The reporting available does not give a timeline for when the correction would actually hit, who collapses first, or how the trillion-dollar capex commitments unwind without dragging the broader market with them.

If you take his post-bubble forecast at face value, the durable winners are the cheap, open-source models capable of running on personal hardware for the unglamorous tasks they are already good at, like transcription and image processing. That is a much smaller and less profitable AI than the one being sold to investors right now, and arguably a more useful one to plan around.