The image is a parrot that repeats phrases it has heard, with dice deciding which one comes next. “Stochastic” means random in a statistical sense. The paper’s point is that a language model learns which words tend to follow which, so it can produce text that sounds coherent while having, in Bender’s words to UW News, “no communicative intent.” Readers supply the meaning and may trust the output more than they should.
You can see the concern in a customer-service bot that confidently quotes a refund policy the company never had. The sentence is fluent and plausible, and it is wrong. That failure is what this site calls an AI hallucination.
The original paper was about more than the metaphor. It also raised the environmental cost of ever larger models, the biases baked into scraped training data and the difficulty of documenting what that data contains. Those are sourced claims from the paper as described by the University of Washington.
Whether today’s large language models are “only” stochastic parrots is still argued. Critics use the phrase to say fluency is not understanding. Others point to models solving new problems as evidence of something more. That is an open debate, not a settled fact. Editorially, the useful takeaway does not depend on who wins it: a model’s confidence is not evidence that it is right, so outputs that matter need grounding in sources and checks. For how next-word prediction works, see How Do Large Language Models Work?.
Sources
- University of Washington News: Large computer language models carry environmental, social risks (March 10, 2021) — University announcement of "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" at the ACM FAccT conference, quoting Emily M. Bender that the models produce seemingly coherent text with no communicative intent; checked October 24, 2026.