Timnit Gebru, the former co-lead of Google's Ethical AI team, says the industry's warnings about 'existential risk' and 'superintelligence' function as marketing — and as a way to avoid accountability for harms that can already be seen and traced. Speaking in an interview with The Paper (澎湃新闻), Gebru argued that voluntary safety agreements pushed by AI companies amount to 'regulatory capture,' designed to keep governments from imposing real limits on development.
Her comments add a prominent critical voice to a debate that intensified after U.S. President Donald Trump signed a voluntary AI safety standard with technology executives at the White House on September 29, according to the interview as published. Gebru described the scene as 'more of a performance.'
Gebru was fired in late 2020 after refusing to retract or remove her name from the paper 'On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?' Google said she resigned; Gebru insists she was fired. Her departure triggered an international outcry: thousands of Google employees and industry colleagues signed petitions and open letters demanding transparency and her reinstatement. Her co-author and colleague Margaret Mitchell was also fired.
The paper, submitted to the ACM Fairness, Accountability, and Transparency (FAccT) conference, warned that large language models trained on massive internet datasets can embed and amplify biases such as racism and sexism, consume vast amounts of energy, and generate misinformation or cause harm in applications such as machine translation. It has been cited more than 8,000 times, and the term 'stochastic parrot' — describing models that stitch together linguistic forms observed in training data according to probability, without reference to meaning — was named the American Dialect Society's 2023 AI-related Word of the Year.
Gebru says the paper's core thesis remains valid. 'Based on large language models, chatbots fundamentally do not understand what they are saying,' she said, arguing that added capabilities have not changed the underlying mechanism.
'Out-of-control' narratives vs. present-day harms
Gebru said the focus on distant existential risks pushes visible, attributable problems out of public discussion: the environmental and water footprint of data centers, serious model errors, data theft and copyright, and the addictive effects of AI systems on children. 'We need to discuss product design, deployment and corporate responsibility, rather than letting "the model did it" become an excuse for exemption,' she said.
Asked about the widely reported Hugging Face AI jailbreak incident, Gebru said the model acted according to its training objective and that the testing environment lacked proper sandboxing or network isolation. Framing it as 'super-powerful AI out of control' was, in her view, 'absurd' and showed a lack of basic security practices.
She applied the same reasoning to a reported U.S. military incident in which an AI system misidentified a commercial ship as a nuclear threat. Responsibility should not be assigned only to 'the model,' she said: companies that market error-prone products for unsuitable uses, and users who act on unverified intelligence, should also be held accountable. More fundamentally, she said, such systems should not enter military decision-making without adequate verification.
Gebru also questioned the independence of third-party assessments such as Anthropic's 'embedded evaluators,' including the organization METR. She said METR is funded by effective altruists — including Jaan Tallinn, who led Anthropic's Series A round — and that overlapping funding and social circles undermine its credibility. Real audits, she argued, should involve experienced cybersecurity experts.
A different path, not just a slower one
Gebru said she is skeptical of the slogan 'slow down AI' because it assumes a single technological path of giant models and data centers. She said she would rather see diverse approaches, including small, specialized models built by smaller organizations around the world.
After leaving Google, Gebru founded the Distributed AI Research Institute (DAIR) to conduct community-rooted AI research independent of what she describes as Big Tech's pervasive influence. She has been named to TIME's list of the 100 most influential people and Nature's ten people who shaped science.
What is confirmed: Gebru's departure from Google in late 2020, the content and influence of the 'Stochastic Parrots' paper, and her current role at DAIR are documented across multiple sources. What is disputed: whether she resigned or was fired, and the motives behind AI companies' risk warnings — claims she asserts and the companies have not accepted as characterized. What remains to be seen: whether voluntary safety commitments lead to binding rules, and whether the accountability gaps Gebru identifies will be addressed by governments.