Cory Doctorow: the 'reverse centaur' who takes the blame for AI's mistakes
A Hearst freelancer got the byline on a 64-page summer reading guide stuffed with books that don't exist — the job of three interns and a fact-checking desk, done by an AI, with him left to absorb the blame. Cory Doctorow calls this a 'reverse centaur': not a human riding a tireless machine, but a machine riding the human.
A Hearst freelancer got a byline on a 64-page summer reading guide full of hallucinated books — work that used to take three interns and a fact-checking desk. Cory Doctorow's name for that setup: a "reverse centaur," where the machine drives and the human absorbs the blame.
On Hacker News, John Nagle, who read Doctorow's full book on the idea, said it undersells how cheaply AI now lets corporations run the kind of surveillance only a government once could afford.
Why a Hearst freelancer became "the AI's accountability sink"
A freelance writer took a byline on a 64-page Hearst summer reading supplement. The list was full of books that don't exist. The job used to require, as Cory Doctorow tells it, "a team of three interns, overseen by an experienced journalist, backstopped by an extensive fact-checking department." Now one person did it, working from an AI's output, on a deadline that made real checking impossible. Doctorow, borrowing a phrase from economist Dan Davies, says the writer's actual job was to be the AI's "accountability sink": paid not to write, but to absorb the blame when the machine got it wrong.
That's his answer to what he calls the AI paradox: how can the same technology be simultaneously miraculous and worthless? The trick, he argues, is in two words: centaur and reverse centaur. A centaur is a human head directing a strong, tireless machine body: the operator stays in charge. A reverse centaur flips it: "a frail and vulnerable person being puppeteered by an uncaring, relentless machine." Same tool, opposite arrangement, opposite result.
Doctorow's own AI use: thirty hours of podcast audio in an afternoon
Doctorow contrasts the Hearst writer's ordeal with his own. Hunting for a quote he half-remembered, he ran Whisper — OpenAI's open transcription model — over thirty hours of podcast audio to find it. He chose the tool, chose the task, and used it only where it made his work better. Nobody handed him an impossible deadline and a mess to clean up afterward. The difference between a centaur and a reverse centaur, in his framing, isn't the software. It's who's driving and who's being driven: the gig worker steering an algorithmic quota versus the cyclist who picks their own route.
What survives when the AI bubble bursts, according to Doctorow
Doctorow calls AI a speculative bubble, and argues that's not a reason to dismiss the underlying tools, because bubbles leave things behind. His reference point is the dot-com crash: when WorldCom went bankrupt, AT&T picked up its abandoned dark fiber for pennies, and that's a meaningful share of the backbone under today's cheap fast internet. He expects something similar once AI funding collapses: "many planned data-centers will never be opened," many existing ones will shutter, but scientists could pick up leftover GPUs "at pennies on the dollar" for climate modeling, and tools like Whisper, already folded into ffmpeg 8.0 for automatic captioning, keep running as commodity software long after the venture money is gone. He draws a sharp line against crypto, which he says will leave behind only "bad Austrian economics and worse monkey JPEGs."
"There is no alternative": Doctorow's answer to it
The essay's title borrows from Margaret Thatcher, whose "there is no alternative" doctrine, in Doctorow's reading, was really an instruction to stop imagining other options. He accuses today's tech CEOs of the same move on a smaller scale: Zuckerberg insisting there's no way to stay in touch with friends without surveillance, Pichai insisting search requires spying, Cook demanding veto power over software, Nadella wanting to monitor keystrokes. Each frames an extractive, reverse-centaur arrangement as the only possible one. Naming that there is an alternative, that the same tool can be built to serve the worker instead of harvest them, is, in his words, "a political act of resistance." The essay previews a full-length version of the argument, The Reverse Centaur's Guide to Life After AI, due from Farrar, Straus and Giroux in 2026.
John Nagle read the book: "not one of his better works"
The thread's sharpest pushback came from John Nagle, a long-time HN presence who'd already finished the book-length version. He wasn't impressed:
"It's all conventional anti-AI wisdom. Too much of it is about the impact of AI on the punditry and writing industries. I expected better from Doctorow." — John Nagle
His complaint wasn't that Doctorow is wrong, but that he stopped short. Nagle wanted the essay to follow its own logic further, into what AI does for corporate power specifically, territory Doctorow has mined in his fiction but skipped here:
"With modern surveillance technology to collect data, and AI to analyze it, it now pays." — John Nagle
His point: the Stasi topped out at roughly 10% of East Germany's population on the payroll because human surveillance is expensive. AI removes that ceiling for companies like Flock and Google, not just for governments.
A neuroscientist's day answers the question the essay leaves open
One commenter supplied the first-hand account Nagle's critique was missing. A developmental autism researcher who runs his own lab's servers, databases, and analysis pipelines put it like this:
"At times it can even seem like it is running the show, and I'm just getting in the way: Reverse Centaur. But other times, I feel like it is helping me leverage my time to explore MY goals much more quickly, and it is empowering." — a developmental autism researcher
For him the deciding factor matches Doctorow's own claim almost exactly: who's setting the agenda, except he found it isn't fixed per job, but shifts hour to hour with the same tool.
The pushback: open-weight models don't need the bubble to survive
Doctorow's bet that a burst bubble leaves useful wreckage behind assumes the current foundation models are the ones worth salvaging. The founder building the infrastructure startup flexenv.com questioned that premise directly:
"I'm not sure his argument holds that foundational models will no longer function after the bubble pops. There's plenty of open weight models that are competitive which are more likely to exist in a world with abundant cheap GPUs." — a founder building infrastructure at flexenv.com
In other words, the residue Doctorow expects, cheap GPUs and orphaned tools, may arrive alongside open models that were never dependent on the bubble's investors in the first place.
Fact-checking Doctorow's own habit, and a defense of it
A search of Doctorow's back catalogue turned into its own mini-reversal. A reader who counted fifteen separate "bubble" warnings scattered through the essay argued the framing is less analysis than a running bit: "He's been pulling the same schtick since 2014 calling many other companies a bubble."
The counter came from the creator of the Luminus web framework, defending Doctorow's track record: "Doctorow has consistently one of the most sober takes on AI in my opinion. He acknowledges that we are in a bubble right now, but also understands that underlying tech has actual merit to it." The thread never resolved which read is right: repetition as pattern-matching, or repetition as consistency. Both sides were arguing from the same fifteen mentions.
Fool's gold: when the centaur looks right but isn't
The most concrete complication of the whole framing came from an engineer who says he's worked in a centaur-style AI setup, pushing back on the tidy binary itself:
"The output is not correct, but in a way that looks correct, and I have to spend hours backtracking…" — an engineer working in a centaur-style AI setup
That's a third failure mode Doctorow's binary doesn't quite name: not a reverse centaur being driven by the machine, and not a centaur steering it cleanly, but a centaur who thinks he's steering while quietly following bad directions.
Doctorow's own essay drew one more complaint about form rather than content: a screen-reader user who counted fourteen raw URLs breaking up the piece said each one gets read aloud letter by letter — "H T T P S colon slash slash" — interrupting the very argument about who technology serves.
Between Nagle's book review, the researcher's split-shift verdict, and the engineer's fool's-gold warning, the thread converges on something sharper than Doctorow's own two categories: the question isn't just who's driving, but whether the person who thinks they're driving can actually tell.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
One essay's fix for AI chaos is giving every worker their own agent
Gavriel Cohen's proof that AI agents are already outpacing human oversight: an engineer who ran dozens of them in parallel and rewrote a legacy codebase in under two weeks. He argues that power needs a name and a face behind it, not a handful of AI labs deciding who gets to use it.