Science & Tech · AI · 5 hrs ago
Researchers warn AI-led research could speed up AI development
A group of more than 20 AI researchers examined whether AI could speed up its own development.
Their paper describes a cycle in which AI writes code and runs experiments to improve models, and those models then help with further research.
The researchers say current AI tools have not yet reached the level needed for an intelligence explosion.
Anthropic data cited in the paper shows AI wrote more than 80% of code approved by its employees in May, though AI still cannot complete entire research projects on its own.
If current progress continues, AI could handle research tasks that take people months by mid-2028, the researchers estimate, while noting that evidence is early and uncertain.
Faster progress could bring benefits such as quicker drug development, but could also accelerate cyberattacks and biological threats.
They call for stronger oversight, including independent auditors inside AI companies, regular reporting and plans to pause automated research when needed.
Researchers warn that automating AI research and development could accelerate progress enough to compress years of advances into months.
They say current AI research automation is not yet at the level needed to trigger an “intelligence explosion.”
The researchers estimate that, if current trends continue, AI could perform research tasks that take people months by mid-2028.
They also warn that faster AI development could bring risks, including misuse and weakened human control.
They recommend measures including company reporting, independent audits and procedures to halt automated research when needed.
- Who
- A group of 22 AI researchers, including Geoffrey Hinton and Yoshua Bengio, examined the issue.
- What
- The researchers warned that automating AI research could speed up AI development and pose risks.
- When
- The article was published on October 23, 2026. The researchers estimated that AI could automate research tasks lasting months by mid-2028 if trends continue.
- Where
- The paper came from the University of Cambridge’s AI Science and Policy Programme.
- Why
- AI could help develop more capable AI systems, which could then contribute to further research and accelerate progress.
This story does not have two clearly opposing sides.
Anthropic’s figures showed AI wrote a low single-digit share of code approved by employees.
The share of research and development tasks completed autonomously by AI was 1%.
Anthropic’s figures showed AI wrote more than 80% of code approved by employees.
The share of research and development tasks completed autonomously by AI had risen to 26%.
Researchers estimated that AI could perform research tasks that take people months, if current trends continue.
- Researchers
- 22 AI researchers
- Paper programme
- University of Cambridge AI Science and Policy Programme
- AI-written approved code
- More than 80% in May 2026, according to Anthropic figures cited in the paper
- Autonomous research and development tasks
- 1% in March 2026; 26% in August 2026
- Researchers’ estimate
- AI could automate research tasks lasting months by mid-2028 if trends continue
What people are saying
A strong oversight system for AI is needed to prevent a sudden intelligence explosion, similar to stationing regulatory inspectors at a nuclear power plant.