Stochastic parrot

Writers and journalists, writer, journalist, author, editor, translator, media, news, Students, student, pupil, university, college, studying, Teachers, teacher, teaching, school, educator, lecturer, professor

Also: stochastic parrots

The claim that language models stitch words together by probability, without any grasp of meaning.

The phrase comes from a 2021 paper by the linguist Emily M. Bender, the AI researcher Timnit Gebru and colleagues. They described a language model as a system for stitching together sequences of words it has seen in its training data, according to probabilities of how they combine, “without any reference to meaning”. “Stochastic” means governed by chance; parrots repeat words they don’t understand.

The paper’s main warnings were about harms that are real today: the environmental and financial cost of ever-bigger models, huge unexamined training data that bakes in bias, and the risk that people trust fluent text that nothing stands behind. Those concerns still apply. The phrase itself took on a life of its own, as shorthand for “it doesn’t really understand anything”, and is often used to argue that these systems are less capable than they seem.

On that narrower question, the evidence has moved. Researchers looking inside models have found internal representations of the world: of a game board the model was never shown, of where places are and when events happened, and of intermediate steps in reasoning, such as “Dallas is in Texas” on the way to “the capital is Austin”. The AI Safety Atlas concludes that models compress what they learn into something like world models, though unevenly. Whether that counts as “understanding” is partly a matter of definition. For safety, what matters is that these systems can do more, and more unexpectedly, than the parrot image suggests.