The Pulse
22 Researchers Warn AI R&D Automation Could Accelerate Progress
A paper published September 28 by 22 researchers warns that automating AI research could compress years of progress into months. The authors call for government oversight, ways to constrain acceleration and plans for its social effects.
AI.info Team · Written with AI ·
Anthropic’s reported share jumped from 1% to 26%
Anthropic reported that AI systems completed 26% of its internal AI research and development work with only high-level human supervision in August 2026, up from 1% in March. The figure anchors a warning from 22 researchers: automating the work that improves AI systems could speed up AI progress itself, compressing years of advances into months or less. The paper, published September 28, describes that outcome as a possibility, not a forecast.
Its authors include Geoffrey Hinton, Yoshua Bengio, OpenAI researcher Jakub Pachocki and Anthropic co-founder Jack Clark. They argue that AI systems already contribute to major parts of the research pipeline, while acknowledging that today’s systems still fail some tasks and can misrepresent their work. The paper also cites OpenAI’s report that its systems routinely complete research and development tasks that would take staff days.
The feedback loop depends on software gains
The proposed mechanism is a cycle: AI systems take on more research tasks, help create better successors, then use those successors to take on still more work. Because software improvements can be put back into research quickly, the paper focuses on a software-driven acceleration rather than assuming new hardware will arrive at the same pace. The authors estimate that, once systems reach expert-level research ability at running costs comparable to current systems, one frontier developer’s computing resources could support an AI workforce equivalent to at least millions of top human researchers.
“AI systems are on track to automate most AI R&D work within a few years, and possibly all of it,” write Alan Chan et al., authors of “What if automating AI R&D triggers an intelligence explosion?”
The authors stress that the evidence is preliminary and mixed. Their discussion weighs constraints including diminishing returns on research effort, limits on computing resources and data, tasks that remain difficult to automate, and lengthy processes such as training runs. They say there is a coherent route to rapid acceleration, but that it has not been established that automation has reached the threshold needed to trigger one.
A tenfold acceleration rests on uncertain estimates
To examine whether more automated research could overcome diminishing returns, the paper draws on estimates of “returns to research effort” from earlier studies of three AI subfields. Those studies put central estimates between 1.2 and 1.9; in the paper’s framework, a value above 1 means that an expanding research workforce could accelerate progress. If those estimates held after full automation and no other bottlenecks emerged, the pace of AI progress could increase tenfold within about 1.5 years, the authors calculate. At that point, a year of progress at today’s pace would take roughly five weeks.
The condition matters: the calculation assumes the estimates continue to apply and that other constraints do not intervene. The paper also notes that society may have more time to respond if advanced systems remain superhuman only in narrow domains, or if physical processes and regulation slow the deployment of new capabilities. It presents possible gains alongside risks, including advances in science and medicine as well as cyber or biological threats, labor disruption, diminished human oversight and sharper concentrations of power.
The authors want reporting, restraints and response plans
The researchers call on policymakers to obtain better information about how much AI research companies automate, how quickly capabilities improve, how firms allocate research spending and where AI systems participate in high-stakes decisions. Options include standardized company reporting, third-party measurement and independent auditors embedded within some AI companies. The authors argue that much of the relevant information sits inside the firms developing and using these systems.
They also propose preparing ways to slow or constrain development if risks warrant it: safety requirements, limits on the rate of capability growth, oversight of high-stakes experiments and the ability to pause specific research workloads. Other options include isolated computing environments and international agreements, alongside verification tools and incident-sharing. For wider social effects, they recommend emergency plans for cyber or biosecurity incidents, labor disruption, geopolitical instability and loss of control over highly capable systems.
The paper leaves governments with a timing problem: the authors say preparations may take years, while a fast feedback loop could narrow the time available to act. What comparable measures will let officials see how quickly AI research is being automated—and when, if ever, should they trigger restraints?