Surprisal Theory Is Tautological Without Rational Grounding — Cotterell Attacks a Foundation of LLM-Based Psycholinguistics

arXiv 2607.21574·medium signal

Ryan Cotterell argues that surprisal theory — the claim that human processing difficulty is an affine function of a token's surprisal under some language model — is unfalsifiable as commonly stated, because the freedom to choose 'some language model' makes the claim tautological absent rational grounding. This targets a large body of work using LLM perplexity as a proxy for human reading difficulty. Anyone citing LLM surprisal as evidence about human cognition now needs to specify which model and why that choice is principled.