Reports published September 28, 2026, described a new paper whose authors warn that automating artificial intelligence research and development (AI R&D) could sharply accelerate progress. They propose oversight measures such as independent evaluations and ways to pause selected projects. The scenario is conditional, not an established outcome.
A paper warns about automated AI research
The paper, What if automating AI R&D triggers an intelligence explosion?, examines what could happen if AI systems take on more of the work used to develop AI. Among the contributors named are Geoffrey Hinton, Yoshua Bengio, Jack Clark and Jakub Pachocki.
The authors describe automated AI R&D as the most likely source of the rapid acceleration they consider. They also outline a conditional scale: if AI systems reached expert-level R&D capability, one developer could direct an AI workforce equivalent to millions of top human researchers. That is a scenario, not a measure of current capacity.
What an “intelligence explosion” would mean
The paper uses “intelligence explosion” to describe a possible surge in AI progress that could compress advances otherwise taking years into months or less. In this account, recursive self-improvement means AI systems helping automate the research and development that could produce more capable AI systems. It does not mean the whole process is already running autonomously.
The authors’ scenarios include cyber or biological threats developing faster than countermeasures, less human control as people play a smaller role in AI R&D, and a country turning a modest lead into a decisive one. They also say AI could help mitigate risks. The consequences remain uncertain.
The oversight measures the authors propose
The recommendations aim to make AI R&D progress more visible and give decision-makers options if development accelerates. They include transparent progress reporting and independent auditors or evaluators, along with measures to constrain the pace of improvement.
The authors also propose preparing to pause selected data-center-based R&D projects, isolating automated R&D systems and developing emergency response plans. These are recommendations, not rules already in force across the industry.
A forecast, with constraints
The authors forecast that, by 2028, AI could fully automate R&D projects that take humans months. This is a projection about a defined kind of project, not a claim that AI can already automate all research. In their assessment reported on September 28, productivity gains from AI R&D automation had not yet reached the threshold associated with an intelligence explosion.
Computing limits, challenges in automating research, diminishing returns, long training runs, supply chains and regulation could all affect the scenario. The forecast describes a possible direction for AI development—not a guaranteed timetable.
