Timnit Gebru, the former co-lead of Google's Ethical AI team who was forced out in late 2020 over a paper on the risks of large language models, says the industry's warnings about runaway AI are increasingly being used to shield companies from accountability for harm their products cause today.
In an interview with The Paper, Gebru said AI firms market their products as potentially uncontrollable "superintelligence" while simultaneously lobbying for self-regulation. The real purpose, she argued, is to divert public attention from the current unreliability of those products and the responsibility that developers and users should bear.
Her comments center on a core question: when AI causes or may cause real-world harm, who is responsible?
From a research dispute to an international outcry
Gebru's dismissal remains one of the most contentious episodes in the AI industry's recent history.
In late 2020, Google asked her to either retract a research paper or remove the names of Google employees from it. Gebru refused to do so without further discussion, and was then told by email that the company was "accepting her resignation" — an account she and many colleagues dispute, saying she was fired.
The paper, "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?", was co-authored by Gebru, linguist Emily Bender, computer scientist Margaret Mitchell and other researchers. It argued that large language models carry significant risks, including environmental and financial costs, training data so vast it cannot be documented or audited, embedded biases, misdirected research effort, and the potential to deceive users because the systems do not understand the concepts behind the text they produce.
The fallout was swift. More than 1,200 Google employees and over 1,500 academic researchers, industry and civil society figures signed letters of protest. The petition described the termination as "an act of retaliation" against Gebru and warned of "research censorship."
Google CEO Sundar Pichai later apologized to staff, writing that the departure had "seeded doubts" and that he was "sorry for that." Gebru rejected the apology, saying it stopped short of acknowledging that the company's actions toward her were wrong.
"It does not say 'I'm sorry for what we did to her and it was wrong,'" she wrote at the time. "I see this as 'I'm sorry for how it played out but I'm not sorry for what we did to her yet.'"
Google maintained that Gebru resigned and that the paper did not meet its bar for publication.
"More of a performance"
Gebru's current critique extends well beyond her own case.
She described a recent White House event at which tech executives signed voluntary safety standards as "more of a performance," arguing that the same leaders who publicly warn of existential risk are seeking self-regulation to avoid external oversight.
She said she had warned from the outset that the "runaway model" and "existential risk" narratives were laying the groundwork for regulatory capture — where regulators meant to protect the public end up serving the interests of the industry they oversee.
Gebru pointed to a reported incident in which a military AI system allegedly suffered a severe "hallucination" during an open-sea patrol, misidentifying a commercial vessel as a threat carrying nuclear material. Officers discovered the error shortly before action was taken, according to her account.
In such cases, she said, responsibility cannot simply be shifted onto "the model." Companies that market products in misleading ways or push error-prone systems into unsuitable uses should be held accountable, as should users who act on unverified AI-generated intelligence, she argued.
She raised a similar concern about consumer products. "If a chatbot says something dangerous to a child, will the company argue, 'That was the chatbot speaking, not me'?" she asked. "We need to discuss product design, deployment, and corporate responsibility, not let 'the model did it on its own' become an excuse for liability."
What "stochastic parrots" actually meant
The phrase from Gebru's paper has since become one of the most widely cited critiques of large language models — and, its co-authors say, one of the most misunderstood.
Bender has clarified that the metaphor was intended to describe large language models that generate text, not all AI systems, and was not meant as an insult. "It was just a description of what these systems actually do," she said.
Gebru says the concept remains relevant. Chatbots based on large language models fundamentally do not understand what they are saying, she argued, and even when output sounds fluent or happens to be factual, it cannot be relied upon as factual because of how the underlying models are built.
She is equally skeptical of claims by some former frontier-lab researchers that models are "self-improving" or developing genuine understanding, saying such claims should be subject to peer review, with published data and code, before being accepted.
What is confirmed, and what is not
The confirmed record includes Gebru's departure from Google after refusing to retract or alter the paper, the scale of the employee and researcher protests, Pichai's apology and Gebru's rejection of it, and the arguments made in the "Stochastic Parrots" paper.
Her assessment of the industry's motives — including her view that existential-risk warnings serve as a liability shield and enable regulatory capture — is her own analysis, not a finding by any court or regulator.
The military AI misidentification incident rests on Gebru's account as reported in the interview; the details have not been independently verified in the supplied reporting.
No regulator has ruled that any AI company used "model autonomy" to evade liability, and the voluntary safety commitments signed at the White House remain commitments rather than enforceable rules.
What remains to be seen is whether governments translate growing concern over AI harms into binding regulation, and whether courts will accept corporate arguments that responsibility lies with automated systems rather than the companies that build and deploy them.