Hinton and Bengio Warn Governments of Possible "Intelligence Explosion" in AI
More than 20 leading researchers and company executives in artificial intelligence have warned governments about the possibility of a so-called "intelligence explosion"—a situation in which AI begins to accelerate its own development so rapidly that technological progress, which usually takes years, could happen in months or even faster.
Among the authors of the new report are Nobel laureate Geoffrey Hinton, one of the pioneers of modern AI Yoshua Bengio, Anthropic co-founder Jack Clark, and OpenAI chief scientist Jakub Pachocki. The Guardian reports.
The report is titled "What if AI research and development automation triggers an intelligence explosion?" The authors consider a scenario in which artificial intelligence reaches the level of leading experts in AI research itself and can independently perform an increasing share of the work in creating the next generations of models.
A key mechanism of such acceleration is called recursive self-improvement. A more powerful model helps develop an even more advanced system, which in turn further accelerates research—and this process begins to reinforce itself.
According to the authors' estimates, if AI reaches expert level in research and development, a single company could theoretically deploy the computational equivalent of the work of millions of highly skilled researchers simultaneously. They consider the automation of research in the field of artificial intelligence itself the most likely source of a potential "intelligence explosion."
The researchers emphasize that such a scenario is not yet an established fact. However, preliminary data, in their view, show that rapid acceleration of AI development through software can no longer be ruled out.
The authors note that artificial intelligence is already participating in the development of its own software. Anthropic claims that AI now generates about 80% of the code within the company, and OpenAI uses autonomous AI agents, particularly in preparing and training new models.
At the same time, the current growth in productivity has not yet reached the level required to trigger self-sustaining acceleration. But the researchers believe that new systems may be approaching that threshold. The report allows that research and development projects that would have taken human teams months could be fully automated by AI as early as 2028.
Authors warn of loss of control and new threats
The report states that a sharp acceleration of AI development could bring major scientific and medical discoveries, but at the same time create new risks. For example, the development of biological weapons or cyberattacks could occur faster than governments and companies can create defenses.
Another risk is related to the gradual exclusion of humans from the development process. The greater the share of research performed by autonomous systems, the fewer opportunities specialists will have for direct control over how new generations of models are created and changed.
Separately, the authors warn of geopolitical consequences. A state or company that gains even a slight advantage in the most powerful AI systems could theoretically quickly turn it into much more significant technological, economic, or military superiority.
At the same time, the researchers themselves note the high uncertainty of such scenarios. Technological progress could slow down due to limitations in computing infrastructure, shortages of specialized materials, the need to build new data centers, and compliance with government regulation. More powerful AI may also accelerate not only the creation of new risks but also the development of ways to defend against them.
Governments urged to prepare emergency plans
The authors of the report believe that waiting for super-powerful systems to appear before starting regulation is dangerous, because once a sharp acceleration of AI capabilities begins, there may be virtually no time left for decision-making.
They propose requiring leading AI companies to regularly disclose information about progress in research automation and allow independent auditors. Another proposal is to predefine mechanisms that allow slowing down development if system capabilities begin to grow too quickly.
The researchers also suggest working with data center operators so that certain AI development projects can be promptly stopped if necessary. Automated research systems, in their view, should be maximally isolated from external networks if there is a risk of escaping human control.
In addition, governments are advised to develop emergency response plans for various scenarios of rapid AI development and to determine in advance which measures should automatically come into force when systems reach a certain level of capability.
Calls for stronger oversight come amid ongoing debate within the technology industry itself. In February, the International AI Safety Report 2026 was published, prepared by more than 100 experts under the leadership of Bengio. It also noted that expert opinions on the likelihood of losing control over powerful systems differ significantly, but some potential consequences could be extremely serious.
In the new report, Hinton, Bengio and other authors do not claim that an "intelligence explosion" is inevitable. Their main thesis is that governments must prepare in advance for such a scenario, because if it occurs, developments could happen much faster than ordinary political and legislative processes.
