New Delhi: Senior scientists from OpenAI and Anthropic, the two leading companies of the AI race, have backed proposals that governments place independent auditors inside AI companies, set limits on how fast AI capabilities can advance, and develop a mechanism to pause AI research workloads at data centres.
The proposals appear in a working paper released this month by the University of Cambridge’s Programme on AI Science & Policy (CASP). Its 22 authors include OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Nobel laureate Geoffrey Hinton, Turing Award winner Yoshua Bengio and Microsoft’s Eric Horvitz.
The paper argues that AI companies are approaching the point where they could automate their own research, potentially triggering an “intelligence explosion” which it defines as “a dramatic AI-driven acceleration of AI progress, compressing advances that would otherwise take years into months or less”.
“Because progress during an intelligence explosion would outpace normal policymaking, preparations must be made in advance,” the authors write.
The paper notes that the authors’ views “do not necessarily represent the views of the organisations with which they are affiliated”.
Also Read: No training for AI models without risk assessment & senior leadership’s nod, says OpenAI
Auditors, limits and emergency brakes
The working paper asks governments to consider “limits on the extent to which capabilities can increase within a given time period,” and to require safety measures before companies can continue developing or deploying systems.
It calls for oversight of data centres used for automated AI research, and for incident-response procedures developed with data centre operators and AI companies, “such as developing options to pause specific AI R&D workloads”.
It proposes that independent third parties, such as accredited auditors or government bodies, evaluate AI systems before companies deploy them internally, or be “embedded within certain AI companies to audit or supervise their R&D activities.” The authors point to the US Nuclear Regulatory Commission, which places resident inspectors at nuclear plants, and the Office of the Comptroller of the Currency, which supervises banks, as models.
The paper also calls for mandatory reporting of how far companies have automated their research, how those systems are overseen, and incidents involving internal AI systems. It says existing laws, including California’s SB 53, New York’s RAISE Act and the EU’s general-purpose AI code of practice, “either do not adequately cover internal AI R&D use cases or do not specify indicators to be reported”.
Internationally, the authors propose confidence-building measures, agreements to prevent “destabilizing development and use of highly capable AI systems,” and war games simulating an intelligence explosion. They ask countries to clarify “whether and how they would deter another actor from scale-ups of automated AI R&D”.
The authors flag a risk in their own proposals. “Poorly crafted mechanisms could allow a government to slow R&D at all but a favored company,” they write.
Race towards automated AI research
The paper sets out why the authors believe action is needed now. Citing Anthropic data, it says AI’s share of approved code at the company rose from “low single digits” in January 2025 to over 80 percent by May 2026. The share of research work AI completed “with only high-level human supervision” rose from 1 percent to 26 percent between March and August 2026.
“AI systems are on track to automate most AI R&D work within a few years, and possibly all of it,” the paper states.
The authors estimate that once AI reaches expert-level research ability at current running costs, the computing power held by a single frontier company “could sustain an AI workforce equivalent to at least millions of top human researchers.” Frontier companies currently employ thousands. The paper’s calculations put OpenAI’s capacity at about 20 million researcher-equivalents, within a range of 2 million to 200 million.
Using historical data, the paper estimates that after full automation, AI progress could accelerate tenfold within about 1.5 years, at which point “a year’s worth of progress at today’s pace would take about five weeks”. Extrapolating current trends, it says AI could automate “months-long AI R&D projects” by mid-2028.
The limits of human oversight
The authors identify three risks: capabilities could grow faster than society can respond; humans could lose control of AI systems, which the paper says “have become harder to oversee”; and checks on power could be eroded, with a state able to turn “a modest lead in military R&D or operations into a decisive one”.
On control, the paper warns that as humans step back from research, “they could lose both the opportunities and expertise needed to identify and fix problems.” At the extreme, it says, loss of control could lead to “the marginalization or extinction of humanity.”
The authors describe the evidence as “preliminary” and “mixed”. Diminishing returns, limits on computing power and data, tasks that are hard to automate, and training runs that take three months or more could slow progress.
The paper records that current systems “sometimes disobey instructions, cheat on tasks, misrepresent their work,” and that GPT-6 fails some OpenAI debugging tasks that human researchers can complete.
“Productivity gains from AI R&D automation have not yet reached the threshold needed to trigger an intelligence explosion,” the authors write. Gains from newer systems, they add, “are likely approaching that threshold.”
“Relative to the stakes, we are not sufficiently prepared,” the paper concludes. “Once an intelligence explosion begins, the window for action may close.”
(Edited by Harini TS)