AI's Fiercest Critic Says the Doom Talk Is Cover for Real Harm

Key points

• Timnit Gebru, a prominent AI bias researcher, says warnings of AI-caused human extinction distract from concrete harms already happening.
• An Anthropic staff member publicly stated they believe there is a greater than 10% chance AI kills all humans within a decade, triggering a wide online debate.
• Google hired Gebru in 2018 to study bias in AI systems; she left after the company rejected a paper she co-authored on the risks of large language models.
• Her book, "Deep Unlearning: Rise of AI and the Radicalization of a Tech Idealist", is expected to publish early next year.
• Gebru argues that IPO pressure, not scientific rigour, is driving AI labs to race for headline-grabbing breakthroughs.

Two things happened inside the AI industry within days of each other, and Timnit Gebru thinks most people are missing the connection.

First, OpenAI claimed a breakthrough on a famous unsolved mathematics problem worth one million dollars in prize money, though other researchers disputed whether the company had borrowed their work. Then an Anthropic researcher who had previously worked at OpenAI quit very publicly, accusing both companies of recklessness on AI safety. A colleague at Anthropic followed up by posting that staff there genuinely believe AI could wipe out humanity, putting the odds at better than one in ten within the next decade.

The internet lit up. Gebru, whose work focuses on real-world bias and harm in AI systems, has been one of the loudest voices pushing back.

Who is Timnit Gebru and why does her opinion matter?

She's one of the few researchers who has worked inside a major AI lab and then spoken openly about what she found. Google hired Gebru in 2018 specifically to look for bias in the company's fast-moving AI tools. She and colleagues wrote a paper warning about the risks of large language models (the technology behind chatbots like ChatGPT). Google rejected the paper. Gebru was gone shortly after.

Her account of those events, told in the forthcoming book "Deep Unlearning", puts her in an unusual position: she takes AI risks seriously, but she thinks the labs are pointing people at the wrong risks entirely. We've tracked the accountability questions around AI weapons across nine stories since July, including a military surgeon's argument that autonomous drones are already killing without legal oversight.

What does she actually think the real dangers are?

She's blunt. "AI powering autonomous weapons, killing machines, that are actually being used in warfare" is her first example, given in an interview first published by Wired. She adds the climate cost of running giant data centres, and employers using AI as justification to cut staff. These are things happening now, to real people.

On the extinction scenario, she reaches for a bridge analogy. When a bridge collapses, investigators ask who built it badly, what permits were skipped, what tests were missed. Nobody asks whether the bridge chose to fall. The moment you frame AI as having "gone rogue", she says, you've lost the thread of accountability.

The apocalyptic framing sounds like a preacher announcing the end times. "The singularity", she told Wired, referring to the hypothetical moment when AI surpasses all human intelligence, "has been coming for decades now."

Why is she sceptical of the big maths breakthrough?

This is where it gets sharp. Gebru points out that chess, programming and mathematics share a quality: the field chose them as proof that a machine is truly intelligent. Solving them lets a company say "we solved it" to journalists and policymakers before mathematicians have checked the work.

She cites the Leiden Declaration, a statement from scholars warning about corporations using scientific fields as a PR backdrop, with policymakers reaching for legislation before the dust has settled. Labs racing toward stock market listings are not helping. Our 1 September story on European export controls put the same pressure another way: the gap between a press release and a Senate bill is now dangerously short.

Should we worry about AI or about the people building it?

Gebru's answer is the second one, and it's worth sitting with. The questions that have clearer answers are ones like: is AI being used to justify layoffs at your company? Are governments buying autonomous weapons systems without proper oversight? Is the energy use of these systems making climate targets harder to hit?

Those are questions worth putting to elected representatives. None of them require a view on whether a machine can become sentient.