Timnit Gebru argues AI doom talk distracts from real harms

WIRED's Backchannel newsletter published a Q&A in which Lauren Goode interviewed AI researcher Timnit Gebru about a turbulent week in the AI industry. Two incidents had dominated the discussion: a fight over a $1 million math problem, including a dispute over whether OpenAI had secretly used other researchers' work to solve it first, and the resignation of an Anthropic researcher, who had previously worked at OpenAI, and who publicly quit his job over how both companies were handling AI safety. In response, another Anthropic technical staffer said colleagues "really do earnestly believe AI could kill all humans," adding as a personal estimate, "I personally think it is >10% within the next decade." The exchanges set off a wider debate over whether such fears are legitimate, overblown, a form of regulatory capture, or self-aggrandizement by the AI labs themselves.

Gebru, introduced in the piece as a prominent technology and AI researcher, was previously best known for her high-profile departure from Google. Hired in 2018 to evaluate bias in the company's AI systems, she and fellow researchers wrote a paper on the dangers of probabilistic large language models; Google rejected it, and she left the company soon after. She has an upcoming book, 'Deep Unlearning: The Rise of AI and the Radicalization of a Tech Idealist,' expected out early next year. In the interview, she says she rejects framings like 'safety and alignment' and argues that people warning of AI doom are trying to distract from the actual harms the tech industry is already causing.

Asked about the math dispute, Gebru questions why AI companies invest so heavily in subjects like programming, chess and math specifically. She says those fields got elevated to stand in for 'intelligence' itself, so a company can claim it 'solved' the field and therefore solved intelligence. She argues OpenAI's preferred story is that its AI achieved something incredible and 'super intelligent,' which is why it targeted the $1 million question, and that if the normal process by which mathematicians vet the significance, novelty and credit of a claim had been allowed to run its course, and everyone had waited for the dust to settle, there would be a clearer picture of whether OpenAI's framing was accurate. She points to something she calls the 'Leiden Declaration' as a warning about what happens when corporations use the field of mathematics this way, and criticizes policymakers for relying on press releases and popular media instead of talking to mathematicians directly. She says the gap between a company's breakthrough claim and a lawmaker such as Bernie Sanders proposing legislation has become very short, offering it as an illustrative comparison rather than describing an actual bill, and calls that gap 'not a good thing.' She contrasts today's 'cutthroat' competitiveness ahead of AI companies' IPOs with her time at Google, when researchers regularly collaborated with counterparts at Microsoft and Amazon, saying researchers today are 'frenetic' because of pending IPOs.

On the Anthropic resignation, Gebru says she has spent her whole career making the same argument from different angles, citing the 'stochastic parrots' paper she wrote with coauthors in 2021 and, further back, OpenAI's earlier claim that GPT-2 was too powerful to release, which she calls overhyped and says amounted to asking the wrong questions. She uses an analogy from her upcoming book: "Can a bridge decide to collapse?" Her point is that when a bridge collapses, nobody asks whether the bridge was ethical or sentient; investigators ask who built it so poorly, and whether the required permits and tests were followed. That, she argues, is the question to ask when an AI system supposedly 'goes rogue': who built it to be flimsy, not whether the system itself decided to misbehave.

Gebru says what is genuinely existential is AI powering autonomous weapons and killing machines that are already being used in warfare, a climate catastrophe that AI infrastructure could make worse, and employers using AI as an excuse to get rid of workers. In her view, the 'machine-god narrative' is meant to distract from those three harms. She compares the AI industry to a mix of cults and true believers, likening it to a preacher predicting the end times, and notes that the 'singularity' has supposedly been arriving for decades.

Asked to walk through how an 'AI extinction' would actually happen, prompted by a social-media post joking about 'an angry GPU' showing up at someone's house, Gebru says she 'can't really envision it.' She says many such scenarios amount to asking a machine to optimize for some task, after which it decides that killing everyone is the way to achieve that goal, though she frames this as how the scenarios 'seem' to work rather than something she believes will happen. She jokes that AI could 'infect your machines,' noting dryly that malware already exists without it. The conversation turns instead to threats framed as real: floods, fires, disruption to the global food supply chain, and the possibility that people could create and use chemical weapons. In an editorial note inserted into the piece, WIRED adds that not long after the interview took place, Anthropic said it had blocked several unspecified attempts by scientists to build biological weapons.

Gebru also raises what she calls a privilege element: someone who isn't worried about things that can actually kill them, such as police brutality, floods or warfare, has the freedom to obsess over the 'machine-god' instead. She closes by questioning why so much attention goes to an idea people have been discussing since the 1940s and 1950s. She describes a game she plays, reading people AI-related quotes she has gathered in her research and asking them to guess which decade each is from. Most people guess wrong, she says, because 'it all sounds the same.'

Key facts

In a WIRED Backchannel interview with Lauren Goode, AI researcher Timnit Gebru says the 'machine-god narrative' around AI extinction is, in her words, 'meant to distract us' from what she calls really existential harms: autonomous weapons used in warfare, climate damage tied to AI infrastructure, and employers using AI as an excuse to cut workers.
She was responding to the same week's dispute over whether OpenAI had secretly used other researchers' work to solve a $1 million math problem first, and to an Anthropic researcher's public resignation (he had previously worked at OpenAI) over how both companies handle AI safety; a separate Anthropic technical staffer said colleagues 'really do earnestly believe AI could kill all humans,' personally estimating '>10% within the next decade.'
Gebru argues that if the normal math-community process for vetting significance, novelty and credit had run its course, there would be a clearer read on whether OpenAI's framing of its math breakthrough was accurate; she cites something she calls the 'Leiden Declaration' as a warning about corporations using mathematics this way, and contrasts today's 'cutthroat,' pre-IPO competitiveness with the cross-company research collaboration she saw during her time at Google.
Google hired Gebru in 2018 to study bias in its AI tools; after she and fellow researchers wrote a paper on the dangers of large language models that Google rejected, she left the company soon after. She points to a 2021 paper on 'stochastic parrots' she wrote with coauthors, and to OpenAI's earlier claim that GPT-2 was too powerful to release, as examples of a pattern she calls asking the wrong questions.
Asked to explain how 'AI extinction' would actually happen, Gebru says she 'can't really envision it': most scenarios she has heard just amount to a machine told to optimize for a goal, then deciding that killing everyone achieves it, an idea she frames as how such scenarios merely 'seem' to work rather than something she believes is realistic.

Why it matters

The interview lands inside the same week's AI-safety flare-up: a dispute over research credit on a $1 million math problem, an Anthropic researcher's public resignation over safety concerns, and a colleague's on-the-record '>10%' estimate of extinction risk. Gebru, one of the industry's most prominent internal critics since her 2018 departure from Google, uses that moment to argue that extinction-risk framing itself works as a distraction, pulling attention and policy energy away from harms she says are already happening: autonomous weapons in active use, AI-linked climate costs, and AI invoked to justify layoffs. That argument competes directly with the safety-and-alignment framing that dominates much of the AI policy conversation, and she explicitly names regulatory capture and self-aggrandizement as possible motives behind the doom talk.

Who it affects

OpenAI and Anthropic, whose research claims and safety culture are both directly challenged in the piece; policymakers, whom Gebru accuses of reacting to press releases instead of consulting mathematicians and other domain experts; the mathematics community, whose normal vetting process she says was bypassed in the $1 million problem dispute; workers facing AI-justified layoffs; and people exposed to the harms she names as actually existential, including populations affected by AI-powered autonomous weapons and by climate impacts tied to AI infrastructure.

How to use it

Read it as a framework for weighing the next AI-breakthrough claim or safety warning, not as a factual dispute to referee. Gebru's implicit checklist: has an independent research community actually vetted a breakthrough's significance and credit before it gets reported as fact; does a company's risk narrative pull attention toward hypothetical extinction and away from concrete, current harms like layoffs, weapons use or environmental cost; and is an individual researcher's personal probability estimate, like the '>10%' figure cited here, being read as a company position rather than one person's opinion.

How solid is it

This is a single interview subject's opinion, and it is presented as such throughout: 'in my opinion,' 'I can't really envision it.' The Google-departure and math-dispute background comes from WIRED's own framing of the week's news rather than from additional sourcing within this piece. WIRED separately notes, in an editorial aside, that Anthropic later said it had blocked 'several' bioweapon-related attempts, without giving a count or further detail. Both the underlying math-credit dispute and the Anthropic researcher's resignation are referenced here as live, unresolved stories, not settled facts.

Risks and caveats

The '>10%' AI-extinction estimate belongs to one unnamed Anthropic technical staffer, offered as a personal view rather than a company position, and neither that staffer nor the resigning researcher is named in this piece. The claim that OpenAI 'surreptitiously borrowed' other researchers' work is presented as a live dispute, not an established finding. Gebru's Bernie Sanders comparison is illustrative rather than a report of an actual bill, and her own extinction-scenario description is explicitly hedged: she says such scenarios only 'seem' to work a certain way and that she 'can't really envision' the mechanism herself.

“This machine-god narrative is, in my opinion, meant to distract us from these.”

— Timnit Gebru