More than 20 AI researchers warn automated AI research could trigger an intelligence explosion

More than 20 AI researchers, including Nobel laureate Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft chief scientific officer Eric Horvitz and Meta's Dawn Song, published a report on 28 September 2026. They warn that automating AI research and development could trigger an "intelligence explosion". All signed in a personal capacity.

What the authors mean by an intelligence explosion

The paper is titled "What if automating AI R&D triggers an intelligence explosion?". It defines the term as a "dramatic AI-driven acceleration of AI progress, compressing advances that would otherwise take years into months or less". The mechanism behind it is recursive self-improvement: AI systems improving their own technology without human intervention.

According to the authors, once AI systems reach expert-level capability at AI research and development, a single developer could direct a workforce equivalent to "millions" of top human researchers. AI systems already write most of the code at the companies building them: Anthropic says AI produces 80% of its own code, and OpenAI uses autonomous agents.

Faster progress, harder oversight

The report states that "preliminary evidence suggests that a software-driven intelligence explosion is possible", which would lead to the "extremely rapid development of highly capable or superhuman AI systems". That could bring medical breakthroughs, but the authors list three risks: biological and cyber threats that outpace countermeasures, a loss of human control as people step out of AI development, and states turning a modest lead, for example in cyberspace, into a decisive one.

The authors stress that these impacts are uncertain: a breakthrough could be delayed by supply chains for special materials or by regulation, and AI could also speed up risk mitigation. Preparing for rapid progress, they argue, "should be an urgent priority, including at the highest levels of government leadership".

Five policy proposals

The paper asks governments to require transparent progress reports on AI-related R&D, including independent auditors embedded inside companies; to limit how fast an AI can improve within a given period; and to isolate automated AI R&D systems.

The authors estimate that research projects which would take humans months could be fully automated by 2028. "Productivity gains from AI R&D automation have not yet reached the threshold needed to trigger an intelligence explosion, but gains from newer systems are likely approaching that threshold," they write. They add: "Once an intelligence explosion begins, the window for action may close."