theatlantic.com web signal

Doctorow tells Galaxy Brain the AI boom can't sustain itself

TL;DR

  • Doctorow argues on the Atlantic's Galaxy Brain that the AI boom's growth assumptions are unsustainable and the bubble's pop will rival the pandemic recession.
  • His centaur vs reverse centaur frame splits 'workers helped by a chosen tool' from 'workers forced to serve a machine's pace,' recasting AI usefulness as a power question.
  • Post-bubble, he expects useful AI to look like cheap open source models running locally for narrow tasks such as transcription.

Cory Doctorow's pitch is not that AI does not work. It is that the way it is sold to your boss is a different question from whether it works for you, and the two have gotten tangled up. He sat with Charlie Warzel on the Atlantic's Galaxy Brain to argue the boom's hype, vision and dreams of endless growth are unsustainable, and that the more honest way to evaluate AI is to ask whether we are using these tools or being used by them.

The frame he keeps coming back to is borrowed from automation theory. A centaur is a person assisted by a machine, the one who chose the tool and gets the benefit. A reverse centaur is a worker forced to serve a machine's pace, fed work by software someone else bought. According to Doctorow, that distinction does a lot of the work people usually try to do with benchmarks. The same model can be genuinely useful when you reached for it and miserable when it was imposed on you.

The harder claim is financial. Doctorow contends AI companies have lured investors through overhyped claims and accounting gimmicks, and that when the correction comes the fallout will rival the pandemic recession. The line that ties the two pieces together is the one he has written elsewhere, that AI may not be able to do your job, but an AI salesman can 100 percent convince your boss to fire you and replace you with an AI that cannot.

The honest caveat is that this is one author with a thesis to sell, on the same week his new book is out. The episode does not put a date on the pop, does not break down how much revenue comes from voluntary use versus mandated rollouts, and assumes a clean landing into cheap open source models running locally for practical tasks like transcription. Each of those is contestable.

What is worth keeping anyway is the diagnostic. Next time a vendor pitches you productivity, the most useful question is not what the model scores on a benchmark. It is whether the people using it picked it up, or had it handed to them with a tempo attached.

Shared on Bluesky by 2 AI experts