One of the AI industry's most persistent critics has restated her case in unusually blunt terms: the debate about artificial intelligence destroying humanity, Timnit Gebru says, is not a scientific dispute at all but a marketing campaign that happens to be aimed at regulators and investors.
Speaking to WIRED's Lauren Goode in an interview published Monday, Gebru — who left Google in 2020 after a dispute over a paper documenting bias and environmental costs in large language models — described the existential-risk narrative not merely as a distraction but as "harmful". Her central argument is about who pays for the warnings. The Future of Life Institute, she notes, was co-founded by MIT physicist Max Tegmark and Jaan Tallinn, the Skype co-founder who led Anthropic's Series A funding round; the same Tallinn funds METR, the evaluator whose report on an OpenAI agent allegedly hacking Hugging Face travelled widely. In Gebru's telling, the funders, the labs and the third-party auditors form a single network, so the public perception of many independent voices saying the same thing is misleading.
Follow the money, she argues, and the message makes sense in a way the stated motive does not. Telling investors that super-powerful machines are coming makes them want in; telling governments the same thing makes them fear missing out; telling regulators that the only risk worth legislating is a fictional superintelligence shifts attention away from data centres, pollution, copyright and labour. "You can always abdicate responsibility," she says — the language of "rogue agents" and "P(doom)" removes culpability from the engineers who built the systems.
The interview lands in a week when that critique has unusually concrete material. Anthropic's IPO prospectus, disclosed this week, combines a $42 billion loss, $518 billion in planned compute spending and a formal warning that its own AI could end humanity — a combination Gebru's account of incentive alignment would predict.
On substance, Gebru also answers the charge that her landmark 2021 paper, "On the Dangers of Stochastic Parrots", has been overtaken by reasoning models. Anthropic co-founder Jack Clark recently called the framing a "memetically fit cognitive virus" that blinded researchers to AI progress. Gebru's reply is that the paper described what large language models are, and that this has not changed: models trained to emit the most probable next tokens do not thereby understand their outputs. She points to work by Samy Bengio's team showing that reasoning benchmark scores degrade sharply under small benchmark modifications — the pattern that indicates memorised format rather than general capability — and to medical and translation failures where fluency hid error, including the automation-bias effect the original paper warned about.
Her governance proposals are deliberately unglamorous: enforce existing deceptive-marketing law; require documentation of training data; address the labour practices of the data workers who label and sometimes impersonate chatbots; stop the unauthorised scraping of data. She argues that these measures, not speculative superintelligence rules, would meaningfully slow an industry whose current market math depends on cheap data and low accountability.
Gebru is not arguing that AI is harmless. Her institute and her forthcoming book — "Deep Unlearning: The Rise of AI and the Radicalization of a Tech Idealist", due early next year — focus on documented harms: bias, environmental load, labour exploitation and the erosion of public trust through fluent but unreliable output. The disagreement is about where the burden of proof belongs.
What to watch: whether regulators take up the boring instruments she names — documentation duties and marketing rules — or continue to legislate around the hypothetical; and whether the IPO-driven war of narratives changes who is treated as an independent expert in the next safety debate.
