What the paper argues
A working paper published by the Cambridge Programme on AI Science & Policy asks what happens if the companies building AI succeed in automating the work of building it. Its title is the question: What if automating AI R&D triggers an intelligence explosion?
An intelligence explosion, as the paper defines it, is “a dramatic AI-driven acceleration of AI progress, compressing advances that would otherwise take years into months or less”. The authors do not claim this will happen. They argue the evidence that it could is now strong enough that preparing afterwards would be too late, because progress during one would outpace normal policymaking.
The signatures are the part that makes this hard to file under advocacy. Twenty-two authors, among them OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft chief scientific officer Eric Horvitz, Dawn Song of UC Berkeley, and the Turing Award winners Geoffrey Hinton, Yoshua Bengio and Andrew Barto.
The evidence they cite
Most of it comes from the labs themselves. The paper cites Anthropic reporting that AI systems’ share of approved code rose from low single digits to over 80% between January 2025 and May 2026, and that the proportion of R&D work completed autonomously with only high-level human supervision went from 1% to 26% between March and August 2026.
OpenAI is quoted saying AI assistance is used “in practically all parts of the company”, with code-executing agents used in training, evaluating and securing future models. Google reports AI is used in “almost all work that involves writing code or configuration, technical design, research ideation”.
The mechanism the paper describes is a loop: AI systems enlarge the effective research workforce, that workforce produces better systems, and those enlarge it again. The frictions are real — compute limits, training runs that take time, tasks that resist automation — but the authors argue software advances feed back almost immediately, where hardware gains wait on years-long construction.
What they want policymakers to do
Three things, in order. Obtain visibility into how much AI R&D is automated inside frontier companies. Develop ways to pace and constrain a scale-up. Prepare society for the impacts.
The specifics are unusually concrete for a paper with industry authors on it. Standardised reporting of R&D automation indicators to governments and third-party auditors. Accredited auditors or government evaluation bodies assessing systems before internal deployment — or embedded inside certain companies to audit or supervise their R&D, on the model of the Nuclear Regulatory Commission and the Office of the Comptroller of the Currency. Limits on how far capabilities may increase within a given period. Oversight of data centres running automated R&D, with prepared options to pause specific workloads. And a requirement that some evaluations run on air-gapped networks, to stop weights leaking and to contain systems that try to escape.
The authors also name the obvious abuse: a badly drafted mechanism could let a government slow research at every company except a favoured one.
What to watch
The paper is a working paper, not a policy proposal with a sponsor. The thing to watch is whether the reporting indicators it sketches turn up in any binding instrument — a national reporting rule, a licence condition, a clause in the next round of EU implementation — and whether the labs whose scientists signed it are willing to report those indicators before they are compelled to.
