Product1 publisher3 min readPublished
Timnit Gebru tells WIRED that OpenAI chose math so it could claim it had solved it
A disputed million-dollar proof and a public resignation at Anthropic landed a day apart. Gebru's answer to WIRED puts the weight on how fast a lab's claim reaches a lawmaker with no way to check it.
The Product Desk · Product desk
What happened
- WIRED reported a major fight this week over a million-dollar math problem and whether OpenAI surreptitiously borrowed other researchers' work to solve it first.
- A day later, an Anthropic researcher who had previously worked at OpenAI quit very publicly, citing concerns about how both companies were handling AI safety.
- Timnit Gebru, whose own work WIRED describes as rooted in AI safety, rejects phrases like "safety and alignment" and says doom warnings distract from harms the industry might actually cause.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Gebru's position is that the math community's vetting of novelty and contribution would have shown whether OpenAI framed the result accurately. A team deciding this quarter has to act before that vetting concludes.
- contradiction One Anthropic staffer puts extinction risk above 10 percent within ten years while a leading critic of the industry says that register of talk exists to distract. Anyone answering an AI risk question in a review has to pick which one to cite.
- precedent On Gebru's account, the gap between a company's breakthrough claim and a Bernie Sanders bill is now short. Rules written that fast bind every team shipping an AI feature, not just the lab that made the claim.
- exposure A roadmap or a sales deck that leans on a lab's headline result inherits the dispute over that result, and the internal owner is the one who has to explain it.
Somebody at your company has to fill in the questionnaire row that asks them to describe the risks of the AI features in this product. This week that person got two pieces of evidence from inside the labs. They point in opposite directions.
Speaking to WIRED's Lauren Goode [20], Timnit Gebru went at the choice of problem before the proof: "why are they heavily investing in solving math problems? Why are they choosing certain disciplines?" she said [9]. Her answer: programming, chess and math were elevated into proxies for intelligence. It is so the labs can say they solved it, she said [10]. The story OpenAI wants you to hear, she said, is "this AI did this incredible thing, and it's super intelligent" [11].
The math community has a whole process for vetting significance, novelty and who contributed. Had the claim gone through it, Gebru said, "we would have gotten a better sense of whether OpenAI was framing this breakthrough accurately" [12]. She pointed to the Leiden Declaration, which warns about corporate use of mathematics, and said policymakers start relying on press releases and popular media when they should be talking to mathematicians themselves [13]. On the pace: "The amount of time it takes now to go from companies claiming an AI breakthrough to Bernie Sanders proposing a bill is really short, and that's not a good thing," she said [14].
That interval is what reaches your roadmap. A rule drafted off a lab's press release lands on the team shipping a summarise button as well as on the lab that made the claim.
WIRED identifies the person who resigned only as an Anthropic researcher who had previously worked at OpenAI, and the colleague who answered him as another technical staffer [21]. That staffer said people at the company "really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade" [3], which is better than one in ten [22]. Gebru answers with a bridge. In her book she writes, "Can a bridge decide to collapse?" [18] and she told WIRED: "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?" [19]
Useful detail for anyone maintaining a vendor review: WIRED reports that Gebru's own work is rooted in AI safety, and that she rejects phrases like "safety and alignment" [8]. That is the vocabulary most questionnaires are written in. The interview is about the public argument, and the only decision-maker in it is a policymaker holding a press release.
Any capability claim you are about to depend on turns on two things: who checked it, by a process independent of the vendor, and whose build it is if the feature fails in your product. The first costs you time, because you will ship after the demo instead of alongside it. Skipping it means you have repeated a claim that, in Gebru's account of the math fight, nobody outside the lab had finished checking [12].
What to watch
- Whether the math community publishes a verdict on novelty and contribution in the million-dollar problem. A verdict would give a better sense of how OpenAI framed the result.
- Whether Anthropic responds publicly to the safety concerns cited by the researcher who resigned, or more staff follow him out.
- Gebru's book Deep Unlearning, expected early next year, for the longer version of the bridge argument.