Timnit Gebru says AI’s extinction warnings obscure present harms

The warning and the business around it

Gebru’s objection is not simply that predictions of human extinction are too dramatic. She argues that the story changes what policymakers and the public pay attention to.

Warnings about a future “god machine” can make current disputes over copyright, data-center pollution and legal accountability seem minor. They can also frame regulation as a choice between trusting the companies to build powerful systems safely and letting geopolitical rivals get there first.

Gebru points to a network she sees behind that framing:

Gebru argues that the appearance of independent agreement can obscure overlapping financial interests. The same billionaires, she says, have backed Anthropic and organizations warning about AI existential risk, as well as third-party groups cited as independent experts. She argues those investors may benefit more than anyone else from an Anthropic IPO.

That, in her view, is the contradiction worth examining: company backers warn that the product could destroy humanity while standing to profit from the company selling it. The warnings also serve different audiences at once—investors hear that they should gain access to powerful systems, governments hear that rivals must not get them first, and regulators are pushed toward hypothetical future dangers rather than present conduct.

Rules for the systems that exist

Gebru sees a familiar pattern in how companies talk about regulation. They ask for international cooperation on existential risks, she says, but resist rules that would hold them responsible for what they do now. In 2023, Sam Altman called for international cooperation; when the European Union passed its AI Act, he threatened to withdraw OpenAI from the bloc.

She points to Lina Khan, the former Federal Trade Commission chair, for a simpler approach: existing laws apply to AI companies too. Gebru highlights three areas where she says enforcement could begin:

Gebru says companies do not even document where their data comes from and will resist being required to do so. She describes hundreds of millions of people around the world doing painstaking data-labeling work. Journalist Lauren Good also raised a recent 404 Media report alleging that a human sometimes stands behind answers from Meta Muse, a new chatbot.

What the “stochastic parrots” argument still says

Gebru co-authored the 2021 paper “On the Dangers of Stochastic Parrots,” which described large language models as systems trained on vast amounts of internet text to produce likely sequences of words. The paper appeared at the Association for Computing Machinery’s 2021 conference on fairness, accountability and transparency, after Gebru’s conflict with Google over the work led to her departure.

The term was a metaphor for how language models work, not a claim that they merely copy exact phrases. Gebru says the underlying point remains relevant: the models predict text from training data, and fluent output can persuade people that a system understands what it is saying.

The paper also raised concerns about environmental costs and the lack of documentation about training data. Gebru says those issues have not gone away. Nor has the risk of people over-trusting automated systems: she cites cases of incorrect medical information and a translation that turned a Palestinian man’s “good morning” into “attack them.”

Jack Clark, an Anthropic co-founder, recently called the stochastic-parrots idea a “memetically successful cognitive virus” that had spread from 2021 to 2025 and led people to underestimate AI. Gebru rejects the claim that the paper is obsolete. It focused on language models, she says, and those remain the basis of chatbots such as Claude and ChatGPT, even if systems also include agents trained with reinforcement learning.

Her larger concern is that claims about reasoning are often stronger than the evidence. Researchers may call a model’s output a “chain of thought,” she says, without knowing whether it reflects actual thought or forms a coherent chain. She also argues that small changes to reasoning tests can make systems fail, and says that without access to training data, evaluation data and code, outside researchers cannot tell whether a model has seen the benchmark before.

Gebru’s position is not that AI systems can never reason. She says she does not know whether they will. Her criticism is about treating names and demonstrations as proof of capabilities, while independent researchers lack the materials needed to reproduce company claims.

I think the central tension in her argument is not between optimism and fear. It is between attention to speculative capability and accountability for systems already deployed. Gebru’s alternative is to support people building different technological futures—ones that do not rely on environmental damage, exploited labor or stolen data. The harder question for AI companies is whether they will accept rules for the products they sell before the promise of a future “god machine” becomes a reason to postpone them.