Leading AI researchers including Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki and Anthropic co-founder Jack Clark warned Monday that an AI intelligence explosion — years of progress compressed into months — could become the most consequential technology shift in history, urging governments to prepare oversight now.
Their paper, What if automating AI R&D triggers an intelligence explosion?, lists more than 20 authors and focuses on recursive self-improvement: AI systems that accelerate their own research and development. The Guardian reported the group says once expert-level AI R&D automation arrives, a single developer could effectively run a workforce equal to millions of top human researchers.
What an AI intelligence explosion would mean
The authors define an intelligence explosion as a dramatic, AI-driven acceleration of AI progress. They argue automated AI R&D is the likeliest trigger because models already help improve successor systems, and new capabilities can be deployed quickly once built. Preliminary evidence, they write, suggests a software-driven explosion is possible — bringing medical and scientific upside alongside extreme control and security risks.
According to the report, summarized by The Guardian and previewed in The Wall Street Journal, risks include biological and cyber threats outrunning defenses, humans losing practical control as people drop out of the R&D loop, and states converting modest tech leads into decisive power advantages.
Who signed the recursive self-improvement warning
Beyond Hinton and Bengio — often called godfathers of modern deep learning — co-authors include OpenAI’s Pachocki, Anthropic’s Clark, and researchers affiliated with major labs and universities. Industry context matters: Anthropic has said AI now writes a large share of its code, and OpenAI has used autonomous agents in model-training work, the paper notes. The full report is hosted by CASP.
Policy steps the authors want before the window closes
The group says once an intelligence explosion begins, the window for action may close. Recommended priorities include transparent AI R&D progress reports with independent auditors inside companies, tools to constrain breakneck development (including datacenter-enabled pauses), isolation of automated R&D systems, and government emergency-response plans. The paper estimates R&D tasks that take humans months could be fully automated by around 2028, with productivity gains nearing — but not yet at — an explosion threshold.
The call echoes other oversight debates covered here, including when Bill Gates pressed Congress on AI regulation and local questions around compute infrastructure in our report on a Boston data center ban proposal. Massachusetts remains a major AI cluster, as outlined in our guide to major artificial intelligence companies in Massachusetts.