Google DeepMind chief executive Demis Hassabis has called for the creation of an independent standards body to test the most powerful artificial intelligence models before they reach the market, and plans to press his case with U.S. policymakers in Washington in the coming days.
July 14 · Google DeepMind
A Referee for Frontier AI: DeepMind Chief Calls for an Independent Testing Body
Demis Hassabis proposes a government-backed, FINRA-style standards organization to vet the most powerful AI models for national-security risks before they reach the market — and heads to Washington to make the case.
30-day
Window before release when labs would voluntarily share frontier models for testing
3
Core risk areas in scope: cyber threats, biological misuse, agentic-AI behavior
1
Capability benchmark threshold — only models above it fall under review
Who falls under the body's remit
Scope is drawn by capability, not company size — thresholds updated periodically.
Frontier-class models
Above the threshold → mandatory review (eventually)
Startups & academia
Below the threshold → exempt
How the testing pipeline would work
1 · Share
→
2 · Test
→
3 · Gate
The proposed body, at a glance
A government-backed public-private standards authority — modeled loosely on FINRA — linked to U.S. National Labs.
Governed by a board of independent experts and open-source representatives
Funded through industry-led contributions
Aims to build a third-party auditing ecosystem
Focus: watermarking, human-readable outputs, and catching deception or guardrail evasion
Early support
Sam Altman called the framework thoughtful ; Elon Musk called it a good starting point — echoing earlier safety arguments from Dario Amodei.
Open questions
Voluntary sharing has no enforcement power . Benchmarks, funding figures and timeline remain unpublished — and rival labs may not submit.
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…