Archive · 29 Sep 2026
AI Researchers Warn Governments About a Possible Intelligence Explosion
Researchers are studying whether AI could eventually help build better versions of itself.
This is called recursive self-improvement.
If it happened, AI progress could become much faster than human research.
The report says years of progress might happen in months or weeks.
Some AI companies already use AI for coding, training models and other research tasks.
However, today’s systems are still far from freely improving themselves.
Faster progress could help medicine and other technologies, but it could also increase cyber, biological and security risks.
The researchers want governments to prepare before rapid acceleration begins.
They also say the outcome is uncertain and could be slowed by manufacturing, supplies, regulation or other limits.
A report says AI systems may increasingly automate research and development of more capable AI.
Researchers warn recursive self-improvement could compress years of progress into months or even weeks.
Anthropic and OpenAI are cited as using AI for substantial coding, model training and other internal tasks.
The report says an intelligence explosion is possible but not inevitable, and current systems have not demonstrated unrestricted self-improvement.
Authors urge governments to monitor AI development, independently scrutinize companies and prepare measures to pause or limit risky work.
- Who
- More than 20 AI researchers and policy figures, including Geoffrey Hinton, Yoshua Bengio, Jack Clark and Jakub Pachocki, as well as leaders associated with OpenAI, Anthropic and Microsoft.
- What
- They warned that automated AI research and possible recursive self-improvement could accelerate AI development beyond the pace at which governments and institutions can respond.
- When
- The report discusses possible acceleration by as early as 2028; it does not provide a publication date in the articles.
- Where
- No specific location is identified.
- Why
- The authors want governments to prepare for a possible intelligence explosion, improve oversight and preserve the ability to pause or slow risky AI development.
Urgent Preparation
Uncertainty and Constraints
How soon acceleration could occur
Urgent Preparation
The researchers warn that once AI reaches expert-level research ability, automated development could rapidly multiply and leave governments little time to act.
Uncertainty and Constraints
The report does not say an intelligence explosion is inevitable; current systems remain limited, and the timing and scale of any acceleration are uncertain.
Need for government intervention
Urgent Preparation
The authors call for monitoring, independent oversight, emergency plans and mechanisms to pause or slow particular AI programs before rapid acceleration begins.
Uncertainty and Constraints
The report notes that regulation, manufacturing, supply chains and access to specialized materials could constrain real-world progress, while the effectiveness of restrictions remains unresolved.
Balance of consequences
Urgent Preparation
Advanced AI could increase cyber and biological risks, enable governments to gain sudden strategic advantages and make human oversight more difficult.
Uncertainty and Constraints
The same systems could also accelerate medical and technological discoveries and might help develop methods to reduce the risks they create.
- Main concept
- An “intelligence explosion” is a hypothetical period in which AI systems increasingly conduct research that helps create more capable AI.
- Potential speed
- The report says years of technological progress could potentially be compressed into months, while another scenario describes a year’s progress occurring in weeks.
- Current evidence
- Anthropic has said AI systems generate a substantial majority of its code, while OpenAI has begun using autonomous agents for tasks including model training.
- Current limitations
- The report says existing systems remain far from unrestricted self-improvement and that current productivity gains have not yet reached the level needed for an intelligence explosion.
- Possible benefits
- Faster AI research could accelerate discoveries in medicine and other technologies.
- Potential risks
- More capable systems could lower barriers to sophisticated cyberattacks and biological threats, strengthen governments against rivals, or reduce human understanding and control.
- Recommended preparation
- The authors propose detailed reporting, independent scrutiny, development monitoring, emergency plans and possible restrictions or pauses on certain research programs.
