Top AI researchers call for oversight of automated model development
Over 1,100 researchers from OpenAI, Anthropic, Google DeepMind, and others urge governments to slow AI progress before recursive self-improvement outpaces human control
More than 1,100 AI researchers, including chief scientists from OpenAI, Anthropic, Google DeepMind, Meta, and NVIDIA, have signed an open letter asking the US government to establish international mechanisms that could deliberately slow or pause AI development. The letter, titled “Pacing the Frontier,” reflects a growing consensus among the people building these systems that the systems might soon be building themselves.
The core worry is recursive self-improvement: AI models contributing to their own research and development in ways that could accelerate capabilities beyond what humans can meaningfully oversee.
The numbers behind the panic
Anthropic disclosed that its flagship model Claude now accounts for roughly 26% of the company’s research and development work. That figure was under 1% just months prior.
The academic community has noticed too. A survey of arXiv papers published between July and September 2026 identified over 1,250 works focused on self-improvement in AI systems.
A separate effort, organized by the AI Evaluator Forum in September 2026, gathered more than 100 experts who signed a letter advocating for independent evaluators to be embedded at every frontier AI lab.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
What the researchers are actually asking for
The “Pacing the Frontier” letter isn’t calling for a permanent halt to AI development. The ask is more measured: give society enough time to build effective governance structures before the technology outruns its guardrails.
The AI Evaluator Forum letter calls for third-party evaluators with genuine independence and access. Some companies, including Anthropic and OpenAI, have already pledged access to external evaluators, but the letter argues these commitments need institutional backing, not just voluntary goodwill.
A UK AI Security Institute report published earlier in 2026 warned that existing oversight foundations would erode without proactive intervention, arguing that the window for building effective governance is closing, not opening.
Why the timing matters
Recursive self-improvement creates a specific kind of problem for regulators. Traditional technology oversight assumes humans are making the key development decisions, and that progress moves at a pace where policy can catch up. If AI systems begin meaningfully accelerating their own capabilities, both of those assumptions break down simultaneously.