“On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” was presented at the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT). Its authors are Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and a fourth author credited as Shmargaret Shmitchell. The title’s “stochastic parrot” is the paper’s central metaphor: a large language model stitches together plausible-sounding sequences of words from patterns in its training data without any understanding of meaning.
The paper lays out several costs of the race to build ever-bigger models. It documents the environmental and financial expense of training, which falls hardest on communities least likely to benefit. It argues that uncurated web-scale training data over-represents dominant viewpoints and encodes racism, sexism, and other biases that the models then reproduce. And it warns that fluent, human-sounding output invites people to attribute intent and trust that is not warranted, with real downstream harms.
The paper is as well known for the controversy around it as for its content. Co-author Timnit Gebru, then co-lead of Google’s Ethical AI team, departed the company in December 2020 in a disputed exit tied to internal review of this paper; the episode became a flashpoint in debates about corporate research independence.
Why a business reader should care: it is a clear, early articulation of risks that became mainstream once generative AI went public, from biased output to overtrust in confident-sounding text. The questions it raises about what training data contains, and who pays the costs, apply to anyone deploying a large model.