In December 2020, Timnit Gebru was dismissed by Google while on leave after refusing to withdraw a research paper on the risks of large language models. The paper was later published in March 2021 at the ACM FAccT conference under the title “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” Five years on, the paper’s 14 pages read less like a warning and more like an early record of problems that are now widely visible.
A research dispute that ended in dismissal
Google had asked for the paper to be retracted or for employee names to be removed. Gebru refused, and the company terminated her employment. The paper had six co-authors. Four were Google employees, and one appeared under the pseudonym “Shmargaret Shmitchell,” later identified as Margaret Mitchell, who was also fired by Google. The paper did not focus on a single product failure. It argued that large language models carry structural risks that are hard to separate from the way they are built and deployed.
It identified five categories: fluent text without understanding, amplification of bias, environmental cost, the inability to audit training data, and language concentration that harms low-resource languages. It also made a broader claim: the financial and competitive incentives facing companies building LLMs make it structurally difficult for safety and ethics concerns to slow product releases. That point was blunt. It was also central.
From hallucinations to biased systems
On the first issue, the paper described what later became widely known as hallucinations. Its argument was that language models assemble outputs by statistical patterning, not by grounding words in meaning. A response can sound coherent and still be false. That gap is now a routine experience for AI users.
The paper also warned that models trained on historical data would reproduce and scale existing bias. One example cited in the article is Amazon’s AI recruiting tool, developed from 2014 and scrapped in 2018 after it was found to disadvantage women applicants. The system had learned from resumes dominated by male candidates and downgraded resumes containing the word “women’s.” A separate 2019 study published in Science found that a widely used healthcare risk algorithm used medical spending as a proxy for illness severity, which led to Black patients being sicker than white patients at the same risk score. After correction, the share of Black patients flagged for extra care would rise from 17.7% to 46.5%.
Emissions and data auditing became visible at scale
Environmental cost was another warning that the paper said was being underestimated. It cited a 2019 study by Strubell and others, often simplified in public discussion into the claim that training one model equals the lifetime emissions of five cars. The article notes that this figure came from an extreme neural architecture search scenario, around 284 metric tons of CO2e, and should not be treated as a general number for every model.
Still, the later trend was hard to ignore. Google’s 2024 environmental report showed that its 2023 greenhouse gas emissions reached about 14.3 million metric tons of CO2e, up 48% from its 2019 baseline. The report tied much of that increase to electricity demand from AI-related data centers, putting pressure on Google’s 2030 carbon neutrality goal.
The paper’s warning on unauditable training data also found a concrete parallel. In December 2023, the Stanford Internet Observatory reported finding 3,226 suspected child sexual abuse material items in the LAION-5B dataset, with 1,008 confirmed by an outside organization. LAION-5B contains 5.85 billion image-text pairs and had been used to train Stable Diffusion. The dataset was taken down after the findings became public. Bigger datasets did not solve the problem. They made blind spots harder to detect.
Low-resource languages and the incentive problem
The paper also argued that English-heavy corpora would deepen language inequality. The article notes that one later distortion of this issue — that “57% of new English web pages are AI-generated” — is inaccurate. What Thompson and others reported in 2024, based on a web corpus of 6.38 billion sentences, was that 57.1% of the sentences belonged to multilingual parallel sets, suggesting low-quality duplicated material likely created by machine translation. The share was especially high in low-resource languages.
That points to a broader deterioration: low-resource languages are not only underserved, they can also be polluted by poor machine-translated text. The deepest claim in Gebru’s paper was not that AI systems would sometimes fail. It was that the surrounding system was built in a way that resists self-correction, with competition outrunning review, scale outrunning auditability, and speed outrunning safety. That mechanism was visible the day she received the email.