On 17 September Anthropic published a number about itself: 26% of its own AI research is now led by Claude, up from 1% in March. Eleven days later that number is the centrepiece of a paper warning where the curve goes, and the chief scientists of three labs signed it.

More than twenty researchers published What if automating AI R&D triggers an intelligence explosion? through the Centre for the Study of Existential Risk at Cambridge. Among the signatories are three Turing award winners, Geoffrey Hinton, Yoshua Bengio and Andrew Barto, with Jakub Pachocki, OpenAI's chief scientist, Eric Horvitz, Microsoft's chief scientific officer, and Jack Clark, an Anthropic co-founder. The evidence is the industry's own disclosure: research completed by AI under high level human supervision went from 1% in March to 26% in August. From there the authors extrapolate, and mark the extrapolation as tentative, that months long research projects could be automated by mid-2028. They warn that unchecked automation risks the marginalisation or extinction of humanity, and ask policymakers for mandatory oversight of the largest labs.

Why this one is different

Open letters about AI risk are a genre, and the genre's weakness is that the people who sign are rarely the people who build. This one is signed by the person running research at OpenAI and a founder of Anthropic, about the specific capability their own companies are shipping, using a number their own companies published. It is also narrower than the genre usually is: not a claim that AI is dangerous, but a claim that automating the research loop is the step that removes the human pace limit, and that this step is measurably underway.

The people who published the number signed the warning about it.

How we got here

1. 12 Sep 2026Amodei asks the industry to pace itself; Altman and Musk agree.
2. 17 Sep 2026Anthropic measures 26% of its own R&D as led by Claude, up from under 1% in February.
3. 22 Sep 2026OpenAI opens safety assessments to outside groups during training.
4. 23 Sep 2026The UN Security Council holds its first briefing devoted to AI safety.
5. 28 Sep 2026Twenty researchers, three Turing winners and three lab scientists sign the intelligence explosion paper.

What it does and does not mean

A working paper is not a finding. The 26% is one company measuring itself, the mid-2028 date is an extrapolation from two points on a curve that the authors themselves flag as tentative, and the word explosion is doing argumentative work that the evidence does not yet carry. Nor does any signatory propose stopping: the ask is oversight, not a halt. What has changed is who is willing to be named. The argument that automating AI research is the dangerous step has moved from outside critics to the people running research inside the labs, which is the same journey the pacing argument made three weeks ago, and it arrived faster.