Source: Demis Hassabis (Substack) — 2026-07-14
Summary
Google DeepMind CEO Demis Hassabis published an essay proposing a U.S.-led, independent Standards Body for frontier AI, modeled on FINRA-style public-private self-regulatory organizations rather than direct government oversight. Under the proposal, frontier labs would voluntarily share models with the body up to 30 days before release for safety testing on dangerous cyber, biological, and "deception" capabilities, funded mostly by industry, with compliance eventually becoming a prerequisite for any model going live to American users. Hassabis wants the body operational before year-end.
Key Takeaways
- The model is explicitly FINRA-like: an industry-funded, self-regulatory body with independent technical experts and open-source representatives on its board, not a direct government regulator.
- Labs would share models up to 30 days pre-release for testing focused on three risk categories: dangerous cyber capabilities, biological capabilities, and "deception" capabilities.
- Funding is expected to come mostly from industry, both to attract world-class technical talent and to fund the compute needed for large-scale safety testing.
- The proposal starts voluntary but Hassabis frames compliance as an eventual prerequisite for any model going live for American users — a soft-to-hard regulatory on-ramp rather than immediate mandatory rules.
- This lands alongside other recent frontier-model policy moves this month (regulatory pressure on open-weight models, an EU/UK-adjacent wave of AI oversight proposals), positioning DeepMind's CEO as pushing a specific, named governance mechanism rather than general caution.