Daily AI CatchupDeepMind CEO Demis Hassabis Proposes a US AI Watchdog to Vet Frontier Models Before Release
Demis Hassabis published a concrete plan for a U.S. body that would safety-test advanced AI before public release, modeled on FINRA. Frontier labs would voluntarily submit models 30 days before release for screening covering deception, bioweapons, and malicious hacking. He wants it operational before year-end, citing that open-source capabilities could enter dangerous territory within 18 months.
Why it matters
💻 Developer · A 30-day pre-release screening window, if adopted, would directly affect how far in advance you can plan around new frontier model releases — worth tracking as a real possibility given Hassabis's standing.
📦 Product · Mandatory pre-release review windows would be a meaningful planning constraint for any roadmap timed around expected frontier model releases — worth factoring in as a scenario.
🎨 Design · No direct design impact — this is AI governance policy.
📈 Business · A funded-by-labs, government-adjacent oversight body is worth scrutinizing for genuine independence — but the proposal itself signals that voluntary pre-release safety review is gaining serious traction among lab leadership, not just regulators.
🤔 Just Curious · The head of Google's AI lab wants to create an official U.S. group that would test powerful new AI models for serious risks — like the ability to help make weapons or hack systems — before they're released to the public.