Mar 2021, paper, University of Washington and Google Ethical AI team researchers

On the Dangers of Stochastic Parrots

A critique of ever-larger language models by Emily Bender, Timnit Gebru and colleagues, published after a dispute over the draft ended with both of Google's Ethical AI co-leads leaving the company.

What it was

"On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" was written by Emily M. Bender, a computational linguist at the University of Washington; Timnit Gebru; Angelina McMillan-Major of the University of Washington; and "Shmargaret Shmitchell" of "The Aether." Its acknowledgments say the paper "represents the work of seven authors, but some were required by their employer to remove their names." It was published in the proceedings of the ACM FAccT conference on March 3, 2021, the conference's opening day, with Gebru listed under Black in AI.

The paper reviewed models from BERT to GPT-3 and the 1.6-trillion-parameter Switch-C and named four risks: environmental and financial cost, training sets too large to document, research effort diverted from understanding, and text fluent enough to fool readers. Its title phrase came from its description of a language model as "a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning."

Read the original

What it changed

The draft became public through the dispute over it. On December 2, 2020, Gebru announced on Twitter that Google had forced her out. Jeff Dean, head of Google AI, wrote in an internal email that he later posted that the paper "didn't meet our bar for publication," and said Gebru had set conditions for staying that Google would not meet. MIT Technology Review, which obtained the draft from Bender, reported on December 4 that more than 1,400 Google staff and 1,900 other supporters had signed a protest letter. On February 19, 2021, Google fired Margaret Mitchell, Gebru's co-lead on the Ethical AI team, after an internal investigation, The Verge reported. Gebru launched the Distributed AI Research Institute (DAIR) in December 2021.

"Stochastic parrot" became shorthand in the argument over chatbots. On December 4, 2022, four days after ChatGPT's release, Sam Altman posted "i am a stochastic parrot, and so r u."

The arguments it moved

Can language models reach general intelligence?

The paper put a name on the claim that fluent output is not evidence of understanding, and the name outlasted the paper. The authors wrote that an LM is "a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning." Altman answered with a joke on December 4, 2022. Geoffrey Hinton has taken the opposite side. Asked by Scott Pelley on 60 Minutes in October 2023, "You believe they can understand?", he answered, "Yes."

Compare every leader's position on this

What are the biggest risks from AI?

The paper framed the risks of large models as present and distributional: carbon cost borne by communities "least likely to benefit," and training data that encodes "hegemonic" views. Its authors later set this against catastrophic-risk arguments. After the Future of Life Institute published its pause letter on March 28, 2023, citing "Stochastic Parrots" among its references, the paper's listed authors replied on March 31: "Those hypothetical risks are the focus of a dangerous ideology called longtermism that ignores the actual harms resulting from the deployment of AI systems today."

Compare every leader's position on this

Positions it bears on

Timnit Gebru, Scale and large language models

Bigger models carry environmental and financial costs, encode the dominant views of whoever is on the internet, and generate text fluent enough to be mistaken for understanding.

Her stance restates the paper's case that bigger models encode the views of whoever dominates the internet.

Critics and counterpoints
Timnit Gebru, Independence of research from Big Tech

Ethics research cannot flourish inside institutions that punish it; the fix is structural, starting with who pays for the science.

Her argument that ethics research cannot flourish inside companies grew out of the dispute over this paper at Google.

Critics and counterpoints
Timnit Gebru, AI hype and anthropomorphism

Chatbots are built to make users believe a mind sits behind the text; she wants researchers and the public to "unlearn" that framing and to judge systems as engineered products with owners.

The paper's warning about fluent text mistaken for meaning is the root of her later argument that chatbots are designed to seem like minds.

Critics and counterpoints
Geoffrey Hinton, Whether language models understand

Insists that large language models genuinely understand, because what they know was extracted from data rather than written by a programmer.

Hinton's view that large language models understand is the position the stochastic-parrot description rejects.

Critics and counterpoints
Emily M. Bender, Language models and understanding

A language model trained only on text learns patterns of linguistic form, not meaning, because meaning is the relation between form and something outside language; when its output makes sense, the sense is supplied by the reader.

Bender coined the title phrase, which applies her 2020 octopus argument that a model trained only on form cannot learn meaning.

Critics and counterpoints
Emily M. Bender, Data and labor

Large models depend on taking other people's writing and art without consent and on poorly paid data workers who label and filter the output.

Bender said in 2026 that the paper's largest omission was exploitative labor, both data workers' conditions and the taking of creative work.

Critics and counterpoints

Sources

Last verified 2026-09-22.