DeepMind CEO Hassabis Calls for US-Led AI Safety Review Body Before Year-End

One of the people building the most powerful AI on earth is now saying it's time to put someone in charge of checking it.

Three key facts

Google DeepMind CEO and Nobel Chemistry laureate Demis Hassabis published a call on July 14 for a US-led AI safety standards body. His proposed template is the Financial Industry Regulatory Authority (FINRA) — a public-private partnership overseen by the federal government but operationally involving industry participants. The body's core function would be pre-release safety testing of frontier AI models, probing for dangerous capabilities in three areas: cyberattacks, biological weapons, and deception. Hassabis explicitly stated he wants this body operational before the end of 2026.

The plan starts voluntary but aims for mandatory review. Under Hassabis's proposal, frontier labs would submit models to the body up to 30 days before release for safety testing. The first phase would be voluntary. Once the testing regime is proven "effective and robust," it would "quickly" transition to mandatory — frontier models that fail the review would be barred from deployment in the US market. The scope covers models developed domestically and abroad, both open-source and closed-source.

This marks a major shift from the industry's "don't over-regulate" stance. Two years ago, the dominant Silicon Valley narrative was that excessive regulation would kill innovation. Now, the person building one of the most advanced AI systems on the planet is calling for an independent body to review model safety — and using the phrase "urgent action" to describe the necessity of addressing artificial general intelligence (AGI) risks. At minimum, this signals that inside Google, AGI is no longer a distant hypothetical discussion but a near-term reality requiring safety guardrails by year-end.

WangDou's Take

Hassabis is playing this brilliantly. Google is getting squeezed in the frontier model race by OpenAI and Anthropic, and now he proposes a 30-day pre-release safety review. Guess who has the deepest bench to staff a compliance team for that? Google. A mandatory review regime inherently disadvantages smaller players and the open-source community. If you're a 20-person startup, how do you navigate a 30-day regulatory review process? Can Meta's Llama series still ship on its usual open-source cadence? The FINRA model sounds fair on paper, but FINRA protects retail investors — who is AI's "retail investor"? Everyone who uses AI. Letting the weapons manufacturers design the arms control treaty has never ended well historically. But it might still be better than letting them operate with no treaty at all.

Source: CNBC · TechCrunch · Axios