AI researchers publish survey on automating R&D

A March 2026 survey paper offers a rare look at how AI researchers themselves talk about one of the field’s most consequential questions: whether AI systems will start helping build their successors. The paper, posted on arXiv as “AI Researchers’ Views on Automating AI R&D and Intelligence Explosions,” is based on interviews with 25 researchers conducted in August and September 2025. Those participants came from Google DeepMind, OpenAI, Anthropic, Meta, Stanford, Princeton and UC Berkeley, among other institutions. The authors — Severin Field, Raymond Douglas and David Krueger — said the interviews found broad agreement that AI systems are getting better at coding and research tasks, but much less agreement on what happens if that trend continues. ### Who was interviewed, and what did the paper actually ask? The paper says 25 researchers were interviewed about automating AI research and development, recursive improvement and the possibility of an “intelligence explosion.” The authors describe the work as a set of semi-structured interviews with people from frontier labs and academia, rather than a poll meant to produce a representative industry-wide estimate. (arxiv.org) August and September 2025 were the interview period, according to the paper. The participant list included researchers affiliated with Google DeepMind, OpenAI, Anthropic, Meta, UC Berkeley, Princeton and Stanford. ### What did the researchers agree on? Twenty of the 25 interviewees identified automating AI research as one of the most severe and urgent AI risks, the paper said. (arxiv.org) The authors also reported convergence on a nearer-term point: participants generally expected AI agents to become more capable at coding, mathematics and eventually AI development itself. The paper says researchers described a progression from systems acting as “assistants” or “tools” toward “autonomous AI developers.” Agreement became thinner after that point. The authors wrote that predictions diverged on timelines, the pace of recursive improvement and the right governance response. ### Where did the biggest disagreements show up? (arxiv.org) The paper reports what it calls an “epistemic divide” between frontier-lab researchers and academic researchers. Academic participants, the authors wrote, were more skeptical about explosive growth scenarios than researchers at frontier labs. Seventeen of the 25 participants expected advanced coding or AI R&D systems to be increasingly kept for internal use at AI companies or governments rather than released publicly, according to the paper. (arxiv.org) The concerns raised most often were concentration of power and faster AI progress happening out of public view. ### What did the paper say about policy responses? (arxiv.org) The paper says participants were split on regulatory “red lines.” By contrast, the authors wrote that almost all interviewees favored transparency-based mitigations. The paper does not present a single forecast or policy program. Instead, it documents disagreement over timelines and over whether some interventions could create new problems of their own. (arxiv.org) ### What about the claim that Claude writes much of Anthropic’s code? A May 20 X thread circulated the paper and linked to underlying data, according to the source briefing provided for this story. (arxiv.org) But I was able to verify directly from arXiv the paper’s publication details, authors, interview sample and headline findings; I was not able to independently confirm from the sources I accessed the specific claim about Claude writing a large portion of Anthropic’s internal code. (arxiv.org) March 5, 2026 is the date of the current arXiv version listed on the abstract page. The paper remains available on arXiv under identifier 2603.03338, where readers can review the abstract and PDF, and the May 20 X thread cited in the briefing points to fuller underlying material. (arxiv.org)