“On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” was published at the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT), with the canonical record at DOI 10.1145/3442188.3445922. The listed authors are Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell (a pseudonym used in the published version for Margaret Mitchell). The ACM Digital Library blocks automated access, so this entry verifies the text against an open-access copy mirrored by the authors’ institutions.
The paper argues that ever-larger language models trained on uncurated web text carry risks that scale with their size. It groups these into several categories: the environmental and financial costs of training, the tendency of the training data to encode and amplify biases against marginalized groups, the inability of the models to understand the meaning behind the text they reproduce, and the potential to use fluent-sounding output to deceive. The “stochastic parrot” metaphor describes a system that stitches together sequences of language it has observed, without reference to meaning, according to probabilistic information about how they combine.
The paper became the most-cited critique of the LLM paradigm and a reference point in debates about whether scale alone is a sound research direction. It is also closely associated with Timnit Gebru’s departure from Google in December 2020, which occurred while the paper was under internal review. Gebru and Mitchell were Google researchers at the time. The specific internal events around that departure were reported largely through journalism and through participants’ own statements on platforms that cannot be live-verified here, so this entry does not reproduce the contested specifics; see the entries on Timnit Gebru and Emily Bender for the parts grounded in primary sources.