Do AI researchers view automating AI research as a severe risk?

The paper's central claim is that leading researchers, in labs and universities alike, treat automating AI research as one of the most serious risks. Of 25 researchers interviewed in August and September 2025, including participants from Google DeepMind, OpenAI, Anthropic, Meta, UC Berkeley, Princeton and Stanford, 20 "identified automating AI research as one of the most severe and urgent AI risks." The interviewees agreed that recursive improvement is possible but disagreed on timelines and governance. The excerpt also reports an "epistemic divide" between frontier companies and academic institutions. Every participant at a frontier company described active engagement with recursive-improvement scenarios and regular internal discussion, which they said is encouraged. Academic participants, by contrast, "often expressed limited consideration of these possibilities."

The reasoning runs through a feedback loop. The introduction describes an intelligence explosion, a scenario I.J. Good introduced in 1966, in which AI systems design successors that can design AI even better, and it cites Yudkowsky's definition in terms of the return on cognitive investment. The milestone it treats as critical is reached when AI systems can do the work of today's AI researchers, because such a system "could improve the development of" its successor. The excerpt states that AI systems "have not yet been able to recursively improve," so the risk is framed as a prospect rather than an observed event. It also notes, without tying it to the argument, that a version of GPT-5 won a gold medal at the International Math Olympiad in August 2025.

The nearest survey note, How soon do AI researchers expect artificial general intelligence?, measures timelines and extinction credence across 2,778 published researchers. This excerpt is narrower, with 25 interviews on one risk, and its disagreement sits on timelines and governance rather than on whether the loop is possible. Its divergence after the autonomous-developer milestone is the open trajectory that What bottlenecks define the path from AGI to superintelligence? says should be tracked through bottlenecks rather than bet on. The feedback-loop premise the excerpt assumes is what Can recursive self-improvement speed up the research process itself? examines: its case, against diminishing returns, is that automation improves artifacts while research efficiency stays fixed, a question this excerpt never addresses. Last, the excerpt reports agreement on "the possibility of recursive improvement" without separating bounded from open-ended forms, the split that Are self-refinement and recursive self-improvement actually the same thing? draws. The excerpt therefore cannot show whether its consensus covers one phenomenon or several.

The excerpt does not give the interview protocol, how the 25 were selected, how "most severe" was coded, or what the governance disagreements were. It is also cut off mid-quote, so the full account of the lab-academia divide is not visible. The causes it offers, "organizational culture, incentives, and proximity to cutting-edge capabilities," are stated as reasons for the divide, not tested. It uses "ASARA" without defining the term. The implication, at the strength the evidence allows, is that the risk is treated as live inside frontier labs and only partly taken up outside them. The 20-of-25 figure counts interviewees who identified the risk; it is not an estimate of how widely researchers hold the view, and the excerpt gives no grounds for judging academic attention as mistaken.