A paper co-authored by Nobel laureate Geoffrey Hinton and computer scientist Yoshua Bengio, together with more than twenty other researchers, warns that an uncontrolled acceleration of artificial-intelligence progress could become the most consequential technological shift in history. The authors, described as godfathers of modern AI, call on governments to act before such an “intelligence explosion” unfolds.
The report defines an intelligence explosion as a rapid, AI-driven compression of advances that would normally take years into months or less. It links this scenario to recursive self-improvement, where AI systems enhance their own capabilities without human input, effectively automating AI research and development. Once an AI reaches expert-level proficiency in R&D, a single developer could command a virtual workforce comparable to millions of top researchers.
The authors argue that automated AI R&D is the most likely trigger for such an explosion because AI already contributes to improving its own technology, and the resulting upgrades can be deployed swiftly. They note that Anthropic reports its systems now generate roughly eighty percent of their own code, while OpenAI employs autonomous agents for tasks including model training. Preliminary evidence, they say, suggests a software-driven explosion is feasible.
To mitigate the threat, the paper proposes three policy priorities. First, it calls for transparent progress reporting on AI-related research, including the placement of independent auditors inside companies. Second, it urges the development of mechanisms to slow down unchecked AI advancement, such as limits on the rate of self-improvement. Third, it recommends preparing emergency response plans and ensuring that automated R&D systems remain isolated from human control.
The report acknowledges that the impacts of an intelligence explosion remain uncertain. While breakthroughs could accelerate medical and technological advances, they might also enable biological or cyber threats that outpace defensive measures. Moreover, reduced human involvement in AI R&D could erode the ability to steer powerful systems, and states could exploit a rapid lead to achieve decisive strategic advantages.
Given the possibility that fully automated AI projects could replace months-long human efforts by 2028, the authors stress that preparing for an intelligence explosion should become an urgent priority at the highest levels of government. They advise coordination with data-centre operators to enable pausing of risky projects, strict isolation of autonomous research tools, and the creation of contingency frameworks for a range of plausible explosion scenarios.