Stochastic parrots

A system trained to predict the next token learns the distribution of the text, and fluency in that distribution is not evidence of understanding what the text is about. The paper's wider argument also covers training cost, the undocumented scale of web corpora, and the bias such corpora carry forward; the term itself names the first point.

Recorded because the collection repeatedly needs to distinguish claims about what models lack from observations of what they do. The knowing-doing gap is the second kind: it is measured behaviour, and it holds whatever one concludes about meaning.