Stochastic parrot co-author: the metaphor is about LLMs, not AI

The phrase "stochastic parrot" has circulated widely since the 2021 paper as shorthand for the idea that large language models chain words by probability rather than grasp meaning, much as a parrot repeats sounds it does not understand. The paper's own framing is narrower than that shorthand: it is about LLMs, the text models trained on vast data through stochastic gradient descent to build statistical representations of language.

Mitchell, one of the co-authors, used a March 2026 blog post to draw the line she says has been blurred. In her account, current AI systems often layer an LLM together with pre-set rules, deterministic programs, algorithms, and machine learning models that do not process language at all. Reading the parrot image as a verdict on AI in general is therefore, in her view, a category error rather than a finding about machine intelligence.

She also warns that words like "understanding" and "reasoning", useful as intuitive metaphors, can harden into terms that get confused with the phenomena they describe. Her proposed fix is precision: when talking about AI, say which system is meant and what it concretely does.