Doctorow: AI is the billionaire fantasy of workless firms

A new Pluralistic essay from Cory Doctorow reframes the enterprise AI story as a fantasy problem rather than a capability one. The line worth stealing is his: "AI can't do your job, but an AI salesman can convince your boss to fire you and replace you with an AI that can't do your job." The essay's animating idea is that what makes AI attractive to a certain kind of executive is not that it works, but that it removes the humans who keep saying no.

Doctorow's central metaphor is Andy Jassy at Amazon, "stuck in the back seat, playing with a Fisher Price steering-wheel toy," with AI cast as the chance to wire that toy "directly into the firm's drive-train, and make movies without writers or actors, factories without workers." The workers in question, in his telling, are the people whose professional ethics interrupt pure profit maximization: nurses, writers, teachers. He goes as far as arguing that hospital operators want "algorithmic death panels to eject sick patients from their beds," which is a framing, not a documented program.

The most concrete AI-specific claim in the piece is that Google's DeepMind said its materials product had advanced the field by 800 years by "discovering 2.2 million structures," and that follow-up scrutiny found those discoveries useless. Doctorow links a chemistry journal paper making that argument, and uses it as evidence that the gap between benchmark claim and downstream utility is a feature of the sales pitch, not a bug.

The honest caveat is that this is polemic, not reporting. Doctorow is arguing a worldview, and a reader looking for a balanced accounting of where AI does deliver real productivity gains will not find one here. What the piece also does not give you is a serious engagement with the internal metrics executives cite when they justify the layoffs, or a case-by-case audit of the companies he names. Take the specific attributions of motive as Doctorow's read, not as evidence entered into a record.

The reason it is worth reading anyway, if you are the person authoring an AI strategy, is that the failure mode he describes is real and testable. If your deployment case rests on removing the people who currently absorb the last-mile judgment work, and the model cannot actually do that judgment work, you have bought the Fisher Price steering wheel. The teams that will look smart in a year are the ones that can name, in advance, which humans they are not replacing, and why.