"On the Dangers of Stochastic Parrots" is published

Emily Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell presented “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” at the ACM FAccT conference. The paper argued that the costs of scaling language models were being systematically under-counted: the environmental and financial burden of training, the tendency of web-scraped corpora to over-represent already dominant voices, the difficulty of auditing datasets too large to document, and the risk of fluent output being mistaken for understanding.

The title supplied the paper’s most durable phrase. A language model, the authors wrote, stitches together sequences it has observed without reference to meaning — a stochastic parrot. Critics of the scaling agenda adopted the term; defenders spent the following years arguing against it.

The circumstances of publication became as widely discussed as its contents. Google had asked Gebru to retract the paper or remove Google-affiliated authors’ names; the dispute ended with her departure from the company in December 2020, characterised by Google as a resignation and by Gebru as a firing. Mitchell, who co-led Google’s ethical AI team, was dismissed in February 2021. Roughly 2,700 Google employees and 4,300 outside researchers signed a letter in support.

The episode set the terms for a long-running argument about whether industry labs could credibly host research critical of their own products, and it hardened a division in the field between researchers focused on present-day harms — bias, labour, environmental cost, concentration of power — and those focused on projected future risks from more capable systems. That division persisted through every subsequent policy debate.