AI leaders warn automating R&D could spark an intelligence explosion

A Cambridge CASP paper published Monday — coauthored by Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, and more than twenty others — warns that automating AI research itself could compress years of progress into months and urges policymakers to get visibility, set constraints, and prepare emergency plans now.

The people building and researching the most capable systems are saying out loud that if AI starts doing most of the AI research work, progress could leap ahead of governments’ ability to see or stop it. They want auditors, speed limits, and emergency plans before that window closes.

On Monday, 28 September 2026, the Cambridge Programme on AI Science & Policy published a research paper, “What if automating AI R&D triggers an intelligence explosion?” CASP is a University of Cambridge program. The paper is Frontier AI Working Paper Series No. 2/2026. The cover says September 2026. R&D means research and development: the work of building the next AI system, not only using one that already exists.

Twenty-two people are named as authors. The paper lists Geoffrey Hinton at the University of Toronto and the Vector Institute, Yoshua Bengio at Mila, the Université de Montréal, and LawZero, and Andrew Barto at the University of Massachusetts Amherst. The executive summary calls all three Turing Award winners, and it calls Hinton a Nobel laureate as well. The summary names Jakub Pachocki as OpenAI’s chief scientist, Eric Horvitz as Microsoft’s chief scientific officer, and Jack Clark as an Anthropic co-founder. The paper’s own lines list Pachocki at OpenAI, Horvitz at Microsoft, Clark at Anthropic, and Dawn Song at the University of California, Berkeley. Alan Chan and Sören Mindermann are the correspondence authors. Christoph Winter is listed with CASP and Cambridge. The paper says the views are the authors’ and do not necessarily represent the organizations they work with. The summary says academic and civil society researchers started and led the effort, and that the summary itself was written by a subset of the authors.

The paper defines an intelligence explosion as a dramatic, AI-driven acceleration of AI progress, compressing advances that would otherwise take years into months or less. The authors call that a break from the fast but mostly steady progress of the last few years. In the conclusion they write: “An intelligence explosion could be the most consequential technological development in human history.” They also write: “Once an intelligence explosion begins, the window for action may close.”

The lab figures in the paper are citations of those companies’ own reports, not a new announcement from the labs on Monday. The paper says Anthropic reported that the share of internal AI research work completed with only high-level human supervision rose from 1% in March 2026 to 26% in August 2026. The executive summary states the same 26% as of August 2026, up from 1% five months earlier. The paper also says Anthropic reported that AI’s share of approved code rose from the low single digits to over 80% between January 2025 and May 2026. The summary says that as of September 2026 OpenAI reported systems routinely completing research tasks that would take staff days.

The authors say preliminary evidence suggests automating this research could add the equivalent of millions of researchers, because an AI system can be copied and many copies can run at once. The paper’s condition is specific. Once systems can do expert-level AI research at a running cost comparable to today’s systems, the computers one leading developer already has could support a workforce equivalent to at least millions of top human researchers. Those systems would help build the next ones, and the authors say that feedback loop could radically accelerate progress. They also say the productivity gains so far have not reached the threshold that would trigger an intelligence explosion, though they think newer systems are approaching it.

The abstract names three extreme risks if that acceleration happens. Capabilities could grow faster than society can adapt. Humanity could lose control of superhuman systems, meaning systems that outperform people. Checks on power within and between countries, companies, and branches of government could be badly eroded. The paper says much of this is still uncertain, and that the stakes are high enough to prepare anyway.

The executive summary asks policymakers for three things. First, visibility into how far AI research has been automated: independent auditors inside the companies, and reports on how much of the research AI is doing, how fast progress is moving, how the research money is split, and whether AI is used in high-stakes decisions. An auditor, here, is an outside checker placed in the company. Second, ways to steer and constrain an explosion: safety requirements, a speed limit on how fast capabilities grow, a way to monitor high-stakes experiments and shut them down, isolated networks so a system cannot leave human control, and international agreements. Third, preparation: emergency plans for major cyber or biological incidents, labor disruption, geopolitical instability, and a loss of control.

The abstract says AI systems are on track to automate most of this research work within a few years, and possibly all of it. The paper does not say a lab has made a binding promise to stop, and it does not say an intelligence explosion has already begun. A mid-2028 line in the paper, about automating research projects that take people months, is presented as a tentative extrapolation of recent trends.

The picture is a crop of the paper’s title page, in a teal and cream frame. The title is visible, and so are author lines that include Hinton, Bengio, Pachocki, Clark, Horvitz, and Song. It is the paper, not a photograph of a press conference.