More than 20 researchers associated with OpenAI, Anthropic, Microsoft and Meta Platforms reportedly urged lawmakers to strengthen oversight of AI systems involved in AI development. The call was reported on September 30. The researchers warned that AI-assisted development could compress years of progress into months, and raised concerns about human oversight, cyberattacks, AI-facilitated pandemics, labor disruption and geopolitical instability.
Researchers urge oversight of AI involved in its own development
The reported recommendations include clearer regulatory oversight of automated research, strict safety requirements and limits on the pace of capability development. They also call for monitoring high-risk experiments and retaining the ability to stop systems immediately.
Other proposals include secure environments for sensitive research, international networks for sharing incident information, clarified deterrence protocols and coordinated agreements to slow development.
AI assistance is not full recursive self-improvement
AI can help with parts of AI development without autonomously building its own successor. Full recursive self-improvement describes a system that designs and develops a successor system on its own.
Anthropic says its systems already contribute substantially to its development work. The company reported that Claude authored more than 80% of the code merged into its codebase as of May 2026. That figure describes Anthropic’s own coding workflow; the full autonomous successor-development loop is a different capability. Anthropic says it has not been reached and is not inevitable.
What human oversight means in practice
A practical question is who chooses the research goal, evaluates the results and can intervene when an experiment is high risk. Anthropic describes human judgment as part of selecting goals and deciding which results matter. The researchers’ reported proposals add monitoring of high-risk experiments and the ability to stop systems immediately.
