Timnit Gebru, the researcher fired from Google in 2020 after co-authoring the stochastic parrot paper, is arguing that the extinction-risk framing coming out of OpenAI and Anthropic this month is a deliberate distraction from concrete harms the industry is already causing. In an interview tied to her forthcoming book Deep Unlearning: The Rise of AI and the Radicalization of a Tech Idealist, shipping early next year, Gebru pointed to autonomous weapons, climate impact, and AI-driven layoffs as the material stakes being drowned out by what she calls a machine-god narrative.
Her intervention follows two events last week that put safety rhetoric back at the center of the AI debate. OpenAI claimed to solve a $1 million math problem, prompting a dispute over whether the company had used unpublished work from academic mathematicians. A day later, an Anthropic researcher who had previously worked at OpenAI publicly quit, citing safety concerns at both companies.
Gebru's read on the math announcement is that discipline choice is doing a lot of the marketing work. She questions why frontier labs pour resources into programming, chess, and math benchmarks, arguing those domains have been elevated as proxies for intelligence itself. She wants policymakers to wait for the mathematics community to vet novelty claims rather than react to press releases, and she cites the Leiden Declaration as a framework for that separation.
“The story that OpenAI wants you to hear is, this AI did this incredible thing, and it's super intelligent.”— Timnit Gebru, AI researcher and former Google Ethical AI co-lead
Key facts
- 01Timnit Gebru argues that extinction rhetoric from OpenAI and [Anthropic](/claude) is designed to distract from concrete harms like autonomous weapons.
- 02An Anthropic technical staffer said internally that AI killing all humans within the next decade is a >10% probability.
- 03Gebru points to a $1 million math problem that [OpenAI](/openai) claimed to solve as an example of manufactured intelligence milestones.
- 04Her upcoming book, Deep Unlearning: The Rise of AI and the Radicalization of a Tech Idealist, ships early next year.
- 05Gebru cites the Leiden Declaration warning policymakers to consult mathematicians directly rather than corporate press releases.
The tempo problem, in her telling, is that the loop from lab announcement to legislation is now measured in days. Bernie Sanders proposing a bill in response to a corporate claim, before the underlying research has been reviewed, is not a functional policy pipeline. Gebru also notes that the collaborative norms she saw earlier in her career — Google researchers working openly with peers at Microsoft and Amazon — have collapsed as OpenAI and Anthropic race toward IPOs.
On the Anthropic resignation, Gebru agrees the labs are behaving recklessly but rejects the framing that recklessness translates into models that will decide to kill everyone. Another Anthropic staffer responded to the departure by writing that people at the company believe AI killing all humans within the next decade is above a 10% probability. Gebru treats that as a category error, not a probability estimate.
“really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.”— Anthropic researcher, Anthropic technical staffer
Her analogy in the book is a bridge. When a bridge collapses, no one investigates whether the bridge was sentient or ethical; they investigate who built it, what permits were skipped, and which tests were not run. Applying that logic to language models, she argues, means holding the companies and their engineers accountable for shipped systems rather than debating whether the systems have agency.
The actual existential threats, in her account, are already deployed. Autonomous weapons are being used in current conflicts. Data centers are exacerbating a climate trajectory that is itself a physical risk. Employers are citing AI as justification for layoffs. Gebru argues these are the harms the extinction discourse is designed to displace, and she frames the industry as a mix of cults and true believers repeating a singularity script that has been running since the 1940s and 1950s.
“When a bridge collapses, you don't analyze whether the bridge was ethical or sentient or why it decided to collapse. You ask, who is the person who built this bridge to be so flimsy?”— Timnit Gebru, AI researcher and former Google Ethical AI co-lead
She is also skeptical of the mechanics of the doom scenarios themselves. Referencing a viral post, she asks whether an angry GPU is supposed to show up at her house. The more coherent scenarios — a misaligned optimizer, malware propagation, biological weapon synthesis — either describe existing threat classes or, in Anthropic's own recent disclosure of blocked bioweapon research attempts, describe risks already being handled as security problems rather than metaphysical ones.
Gebru's critique will not land cleanly. She is arguing simultaneously that frontier labs are reckless and that their loudest safety claims are marketing, a position that irritates both the accelerationist camp and the existential-risk camp. It also downplays the possibility that some Anthropic researchers genuinely hold the beliefs they state publicly, rather than deploying them strategically. The recent wave of departures from frontier labs, which AI Chat Daily has covered over the past two weeks, suggests at least some of the internal alarm is sincere.
The market implication is that the AI safety discourse is fragmenting into at least three camps — extinction-focused labs, harm-focused academics like Gebru, and regulators trying to synthesize both — and each camp is now writing legislation. That fragmentation matters because it determines which harms get priced into compliance costs and which get ignored. If Gebru's frame wins, expect more scrutiny of defense contracts, labor disclosures, and data center siting. If the extinction frame wins, expect licensing regimes and compute governance. The two agendas are not additive, and the companies funding both know it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
