The Claude Delusion

Doctorow turns the word “hallucination” around. Model errors, he says, are better called errors or defects. The hallucination that matters is ours: when we use a chatbot, “we hallucinate the person on the other side of the interaction” (Doctorow). People reflexively read intention into anything that looks designed, and chatbot output sets off that reflex with nothing behind it. He borrows Mark Fisher’s definition of the eerie, something present where there should be nothing, which comes from Fisher’s 2016 book The Weird and the Eerie.

His analogy is a sunset. To a believer it is a chosen work; to an atheist it is striking but chosen by no one. Chatbot text, produced by statistics over the frequency of human words, is the second kind. He argues that familiarity wears the effect down: repeated use exposes a characteristic smoothness, and experts see it soonest because they know the grain of their subject. A good writing teacher, on this account, is someone who finds the non-standard texture in a student’s work, which is exactly what a statistical average smooths away.

He names two obstacles to what he calls becoming an “AI atheist”. Most people lack the technical grounding to picture a system that acts without intent, and some billionaires, who in his reading already deny full personhood to most people, see chatbots as their equal. He closes by setting AI personhood against the rights-for-nature movement, casting the former as a repeat of corporate personhood.

Two points of fact. The essay asserts that most people’s hallucinations fade with exposure, but offers no data for it. And despite the title, it never names Claude or Anthropic: the argument is about chatbots in general, and coverage that frames it as being about Claude (AI Weekly, 23 September 2026) reads that in from the headline.