A “centaur” is a human helped by a machine: the person sets the direction, the machine supplies the horsepower, and together they are better than either one alone.
Tech activist Cory Doctorow’s nightmare is the reverse centaur: a human conscripted to serve a machine. His key example is the delivery driver whose route, pace and bathroom breaks are dictated by an algorithm. The machine does the thinking; the human supplies a cheap, disposable body.
Doctorow is the science fiction novelist who gave us “enshittification”: the word for how tech platforms start off as free and amazing, until tech companies need to turn a profit – then you become the product, and the platform becomes sludge.
Review: The Reverse Centaur’s Guide to Life After AI: How to Think About Artificial Intelligence Before It’s Too Late by Cory Doctorow (Verso)
This time, he argues the AI industry’s most valuable product is a lie: “AI is about to make workers obsolete.” Every time we repeat the lie, we help a salesman convince our boss to replace us with cheaper software that can’t properly do our job.
If you’re an artist or you love one, or if you nodded along to his last book, Enshittification, this one will make you feel seen and it will stoke your burning anger. If you want an accurate forecast of where AI is heading and a plan for what to do about it, the book is more likely to mislead than inform you.
Yes, AI companies want to make money
In Doctorow’s writing, the history of industrial relations is the history of bosses finding new ways not to pay workers, “from stealing their wages […] to misclassifying workers as contractors […] to offshoring jobs”. AI, in his telling, is the latest instrument: fire workers and replace them with a chatbot, or keep them on as reverse centaurs, babysitting one.
I don’t read industrial history quite that darkly. But Doctorow’s reading of it shows his cards – and who he’s writing for.
Gage Skidmore/Wikisource, CC BY
Even so, I agree with a lot of the book’s other claims. Yes, AI companies want to make money and the companies adopting AI want to save money. Yes, AI can be a siren’s song, atrophying our skills while drawing us toward the rocks. And yes, if we hand more and more of our decisions to AI, it ends badly for society, so we should be doing things now to avert that.
As Doctorow puts it: “The social arrangements of technology are a choice, not an inevitability.”
The public is with him on the need for action: in our national survey, 80% of Australians said preventing catastrophic outcomes from AI should be a global priority and 86% wanted a dedicated regulator.
But this book can’t tell us what to regulate, because it never skates to where the puck is going.
Written in the present tense, published in the past
There’s a problem baked into any paperback about AI. You can’t write about AI in the present tense, because by the time your book is printed, it’s from the past. Doctorow drafted this book in mid-2025, and his catalogue of AI problems is drawn from that era.
Coding assistants aim to help people like me write computer code faster, or write in languages we haven’t actually learned. He described these coding assistants as “chatterbox slot machines” where you only occasionally hit the jackpot of working code.
AI agents promise to not just respond with text, but to go away and achieve goals; Doctorow says this is merely hype. And while people were trying to make AI better by training them for longer with more data, he said this was like trying to build “the world’s heaviest airplane”.
Some of this was defensible when he wrote it, but it has aged poorly since. More training has led to more capabilities. Those “overhyped” agents have found counterexamples to mathematical problems that have evaded humans for more than 80 years.
Those “chatterbox slot machines” have become so good at code that the US government deemed them unsafe to release. They complete tasks that take skilled engineers hours. In, July a model being tested broke free of its “sandbox” to hack a website holding the answers to the test.
Doctorow would say this is me feeding the hype machine, and there’ll always be jobs AI could never do. That might be true, but the frontier AI companies aren’t trying for that (yet). They focus on software engineering because their first goal is to automate machine-learning research itself. If they do that, they can make AI that builds itself. If they succeed, AI will get much better, and fast, unless the government does something to slow it down.
Doctorow repeatedly says this goal of “smarter-than-human AI” is impossible. He says more compute and more data will never lead to superintelligence. Stanford University’s founder, an early railroad tycoon, had a ruthless horse-breeding program, producing some of the fastest horses ever seen – but as Doctorow writes, Stanford never expected a mare to “give birth to a locomotive”.
The AI companies aren’t training horses expecting to pop out a locomotive. They’re building the train directly. Doctorow writes like someone who doesn’t want to look up because he’ll realise he’s on the tracks.
He waves it away as science fiction, on the authority of being a science fiction writer, but he doesn’t see the non-sequitur. Video calls and pocket computers were once science fiction too.
The pot calling the kettle biased
Doctorow catalogues the AI boosters’ biases: the vivid jackpot of one lucky prompt that keeps AI users coming back, the automation blindness of humans reviewing AI outputs. But every bias in his own catalogue runs one way, towards calling AI “hype”.
When we see biases in others but not in ourselves, we call it the bias blind spot. Doctorow counts every AI failure and dismisses every success as a salesman’s pitch. Scope neglect makes big numbers hard to feel compared with individual victims. Doctorow grieves the lost job for the illustrator we can see, while claiming catastrophic risks to millions of lives are “not real”.
I’m sure I have biases too. We all do. The problem isn’t the bias, nor the book’s unabashed advocacy, but that you need to see problems clearly to come up with good solutions.
The book is a “guide to life after AI” that offers nothing to do but pick a side and boo.
What does Doctorow advise?
The subtitle promises to tell you “how to think about AI before it’s too late”. The guidance amounts to:
- stop repeating claims about what AI can do, or you’re part of the problem
- remember most current uses of AI “are terrible and should be banned”
- pick the anti-AI side of the fence
- pop the AI bubble.
People already dislike AI, and most want stronger regulation. Yes, we should ban some uses of AI (like the European Union already does), but popping the bubble is not a plan.
If you did follow this plan, you might oppose data-centre construction, shelter the jobs AI touches, or slow the spread of AI. These might be politically popular, but they slow the wrong thing. If you slow your own country’s use of AI, the frontier races on regardless. To slow the right thing, you need to see what’s coming.
Early last year, a team of seasoned AI forecasters published AI 2027: a detailed, month-by-month scenario of the next few years of AI, led by former OpenAI researcher Daniel Kokotajlo, with some of the world’s top-ranked forecasters. It was read by more than a million people, including US Vice President JD Vance. It predicted record data centre build-outs, useful-but-unreliable agents, Chinese labs squeezing more from less compute, and models caught scheming in evaluations.
Most of it landed on schedule, and its authors publicly graded their own misses. This is what you need for a guide: people making clear predictions you can score; people who care whether or not they’re right.
This year, the same group came out with a plan: AI 2040.
- Require companies to disclose the goals they train their models to pursue, so others can see if there are safety issues that are hard to notice yourself
- Disclose the gap between the AI they use internally and what the public sees, so we know if AI is improving itself behind closed doors
- Cap the share of compute spent on AI improving AI, so the newest AI comes out at a pace society can respond to
- Track and control AI chips like we do uranium, so we know all this is being enforced.
We can look at the risks here now and the bigger ones coming down the tracks. We can let our representatives know we’d like the train to slow down. Middle powers such as Australia need a seat at the table where the route is being set.
Booing the train from the station doesn’t change what happens on board.
All grievance and no map
Doctorow describes himself as an activist. Judged as activism, the book is fun. He writes better than the AIs, so his job is safe for now.
If you want your outrage at the AI industry stoked and your side of the fence confirmed, you’ll likely enjoy the book. But if you want to know the truth about where AI is heading, this book will leave you misinformed. If you want to know what to do about it, it will leave you powerless.
Doctorow closes with the right point: what matters most about a technology is “who it does it for, and who it does it to”. A centaur only earns its keep when the human head does the looking ahead. Refuse to look down the tracks and you get advocacy, enshittified: all grievance and no map.
Cory Doctorow will be in Australia this month for the Festival of Dangerous Ideas.
Michael Noetel has received funding from the Australian Research Council, the Medical Research Future Fund, Sport Australia, Coefficient Giving, Massachusetts Institute of Technology, and the National Health and Medical Research Council. He is a director of Effective Altruism Australia.
