DeepMind CEO pushes FINRA-style AI regulator to end ad hoc reviews

Demis Hassabis wants to borrow a page from Wall Street’s playbook. The Google DeepMind CEO floated a proposal Tuesday for a new AI standards body modeled after FINRA, the financial industry’s self-regulatory organization. The pitch is straightforward: frontier labs would voluntarily hand over models for review up to 30 days before release. If the process proves itself, it flips from voluntary to mandatory for any model hitting the U.S. market. Labs would also be on the hook to work with the body on post-release vulnerabilities.

The timing isn’t accidental. Hassabis is drawing a sharp contrast with the status quo — those ad hoc government reviews of Anthropic’s Mythos and OpenAI’s Sol that got hammered for lacking technical chops and making decisions behind closed doors. His alternative puts industry experts in the driver’s seat, funded by the AI labs themselves but operating independently. It’s the kind of structure that might actually survive the current political climate, where White House AI advisor Sriram Krishnan has already shot down anything resembling an FDA for AI.

Staffing is where this gets interesting. Hassabis envisions open source representatives and technical experts from within the industry running the show, with the option to farm out evaluations to the growing ecosystem of AI safety groups that specialize in specific risks. The financial backing from labs would, in theory, let them pay enough to actually attract and retain people who know what they’re doing — something the government’s reviews conspicuously failed at.

Whether labs actually play ball is the open question. Voluntary compliance has a mixed track record, and the jump to mandatory oversight requires a political consensus that doesn’t currently exist. But Hassabis is framing this as the pragmatic middle path: technically focused, innovation-friendly, and designed to ratchet up if things get serious. “It is designed to keep up with the field’s acceleration,” he writes — a nod to the uncomfortable reality that regulators are always playing catch-up.