The slogan ‘stochastic parrots’ outgrew the argument that coined it

Ten pages of citations, zero experiments

“On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” appeared at the 2021 ACM Conference on Fairness, Accountability, and Transparency. Its authors are Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell, the last credited as “Shmitchell” because Google's legal team objected to her name appearing before the paper cleared internal review.¹ It runs ten pages, citing 157 sources. It became one of the field's most argued-over documents partly because of what followed: Google asked Gebru to retract it or drop her name; she refused. She was fired. The dispute made national news before most of the field had read it.²

The paper itself reports no dataset its authors built, no model they trained, no test they ran. Every number in it belongs to something else the authors read. That is not a flaw; FAccT accepts literature reviews alongside experimental work. It changes what citing the paper can support: citing it for a fact about model behavior cites an argument built from other people's measurements.

The paper's title asks whether models can be “too big.” It organizes its case as four risks: cost, undocumented and skewed data, misdirected research effort, and fluent text produced without meaning it, which the paper says enables harm downstream. Yoav Goldberg, a computational linguist and not a co-author, flagged a problem with that frame: none of the four risks is about size.³ A small model trained on the same skewed text carries the same documentation debt and bias, and can be as fluent and as ungrounded in meaning. “Big” named what alarmed the field in 2020, months after GPT-3 shipped. It is not a condition the paper's argument requires.

The carbon math belongs to someone else's meter

The first risk carries the most measurement behind it. None of the measuring was done by this paper's authors. They cite Emma Strubell's 2019 accounting of the compute cost behind large NLP models. Training one version of BERT released roughly as much carbon as a round-trip flight across the United States. Training a Transformer model tuned by neural architecture search released about 284 tons of CO₂, comparable to five cars driven for their working lives.¹ Strubell's team priced that same search's marginal gains: squeezing one more tenth of a BLEU point, a standard machine-translation score, out of a translation model cost about $150,000 in additional compute.¹ A separate estimate the paper cites, from outside Strubell's group: the compute used to train the largest models grew 300,000-fold between 2012 and 2018, outpacing hardware's two-year doubling.¹

None of those three figures belongs to Bender, Gebru, McMillan-Major, or Mitchell, and not all three belong to Strubell's team either. The paper's contribution is synthesis: it gathers outside accounting, Strubell's and a separate 2018 compute estimate, and argues the field should weigh that cost before training the next model, not after. The argument is only as strong as those numbers, a strength: they are measurable. The paper cites. It does not calculate.

The 272,000 documents nobody read

The second risk works the same way: real figures borrowed from elsewhere. Training text for language models comes from scraped internet content, filtered by automated rules, not read by anyone. The paper cites a separate study's audit of GPT-2's training data, the corpus that indirectly built GPT-3's: it still carried 272,000 documents from unreliable news sites and 63,000 from subreddits later banned.¹ Internet access skews toward whoever has the time, money, and freedom to be online, and the resulting text skews too. The paper cites survey data putting Reddit's users at 67 percent male, and separate research putting Wikipedia's editors at 8.8 to 15 percent women, depending on language edition.¹

The fix the paper proposes existed. Data Statements for NLP, by Bender and Batya Friedman, and Datasheets for Datasets, by Gebru and six coauthors, proposed the same practice three years earlier. Both call for a training set to ship with a spec sheet: where the data came from, who isn't represented, what it should not be used for.⁴⁵ “Stochastic Parrots” did not invent data documentation. It made the case, with the 272,000-document figure attached, for why skipping documentation grows costlier as datasets grow too large to audit.

Goldberg's second objection lands here: the paper frames its curation preference as a neutral fix, when it is a value choice. It wants a model to reflect “the world as we believe it should be,” not the internet's raw distribution.³ A model could reflect language as used, biases included; the paper never engages it. The documentation call stays empirically grounded, the curation choice on top of it asserted.

Where the citations run out

The paper's other two risks lack a Strubell study or a demographic survey. The third says chasing scale pulls funding from other work, a claim about incentives, argued rather than measured. No citation counts how many papers, grants, or hires went to scaling versus alternatives before 2021. It is a reasonable reading of where compute flowed. It is not a number.

The fourth risk produced the phrase “stochastic parrot,” and carries the least independent evidence of the four: philosophical, not empirical.

The paper's citation for that claim is not a dataset or a benchmark. It is Bender's own earlier paper with Alexander Koller, “Climbing towards NLU”: a system trained only on the form of language has no way to learn what that language refers to.⁶ The stochastic-parrot risk rests on that argument, as do the harms it predicts: a fluent system generating misinformation or biased text with nothing to check it against.

Carbon accounting is arithmetic once Strubell's numbers exist. Whether a system that predicts text has “understanding” is a dispute philosophers have not settled either, and the paper takes a clear side.

What five years did to each risk

Environmental cost held up, and got worse. Stanford's AI Index Report puts GPT-4's training at $78 million and Gemini Ultra's at $191 million, using cloud-rental compute prices.⁷ The same analysts' later paper, using amortized hardware costs instead, puts the same two runs lower: $40 million and $30 million.⁸ The exact figure depends on the accounting method. Costs rose by orders of magnitude either way. Emissions from training BLOOM, GPT-3, and OPT ran 25 to 500 tons of CO₂, against roughly 4 tons for one person's annual footprint.⁹¹⁰¹¹ The paper's warning was not heeded. It was priced in.

Documentation became routine practice, unevenly. Data statements and datasheets are standard requests at major NLP venues, closing the gap the paper named. Adoption is thinnest for the largest training runs, the paper's complaint about datasets too big to audit.

The research-priorities claim is hardest to grade: never stated as something evidence could confirm or deny. Compute since 2021 went toward larger runs, not the interpretability work the paper wanted prioritized, consistent with its worry but not a measurement of it.

The meaning argument is the one 2021 could not settle, and 2026 still has not. OpenAI's o1, released September 2024, scored 83.3 percent on a set of competition math problems where GPT-4o scored 13.4 percent, and 89.0 percent on competitive-programming problems where GPT-4o scored 11.0 percent.¹² Whether that gap is understanding, or more powerful pattern-matching than expected, is contested inside the same research community that documented it. Some attribute it to in-context learning, not independent reasoning.¹² The harm side is not contested. PII extraction attacks are demonstrated at production scale: a 2023 attack, catalogued in a 2025 survey of the field, extracted memorized personal information in 16.9 percent of 15,000 attempted queries, 85.8 percent of it genuine.¹³ The paper's speculative harm became measured. The argument it rested on stayed unsettled.

“Stochastic parrots” circulates as a verdict on capability: this model does not reason, so dismiss what it did. The paper's claim, from Climbing towards NLU, was narrower: a system trained only on form has no way to learn what its words refer to, a point about grounding and intent. The paper never tested whether scale would produce reasoning-shaped output, only that such output, however sophisticated, still wouldn't be meaning in its philosophy's sense. Five years of benchmarks have not settled whether that definition is right, and have made the paper's original claim harder to state as fact, easier to borrow as something bigger.