Google DeepMind chief executive Demis Hassabis is calling for a new, U.S.-led standards body to police the world's most powerful artificial intelligence systems, modeled on the securities industry's self-regulatory organization FINRA and answerable to the American government.
A Framework for Frontier AI · Demis Hassabis
A "FINRA for AI": a U.S.-led watchdog to police the world's most powerful models
Google DeepMind's chief executive wants an industry-funded, government-accountable body — modeled on the securities self-regulator — to run pre-release safety tests on frontier AI, and wants it operational within months.
30days
Pre-release window labs share models for testing
2026
Target for the body to be operational — by year's end
Q
"Frontier" benchmarks updated quarterly to define scope
Two-phase rollout — voluntary now, mandatory later
PHASE 1 · VOLUNTARY
Labs share frontier models for up to 30 days before release for safety testing.
→
PHASE 2 · MANDATORY
Once proven, testing is tied to entry into the U.S. market — required for access.
What the watchdog would test for
Cybersecurity threats
Biological risks
Nuclear-related capability
Signs of deception
Agentic planning & self-improvement
Guardrail bypass
Watermarking
Reasoning legibility
Applies to any model clearing frontier benchmarks — open or closed, regardless of developer nationality. Industry-funded, staffed by technical experts, accountable to the U.S. government.
Supporters see
Scientific credibility and alignment with other lab leaders
A workable bridge toward international AI standards
A coherent answer to ad hoc export-control friction
Skeptics warn
Industry funding risks regulatory capture
Constitutional doubts over delegating power to a private body
Unclear leverage over Chinese and open-source developers
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…