Demis Hassabis proposes a FINRA-like body for frontier AI: industry-funded, operating under federal oversight, and able to set common safety expectations. The analogy matters because FINRA combines specialist rulemaking with supervision beyond ordinary voluntary coordination.
One practical starting point is controlled model sharing. Developers could provide access to frontier models for structured testing, incident reporting and evaluation against common protocols. A scheme that begins as self-regulation can then move from voluntary participation toward certification, with the certificate signalling that a model has met defined safety and governance checks.
The safety value is clear. Shared methods can reduce duplicated testing, make failure reporting more comparable and give buyers a legible basis for assessing safeguards. A federal backstop can also make standards less dependent on the goodwill of any one company.
But certification can become a barrier as easily as a safeguard. If the largest labs write expensive, opaque requirements around their own capabilities, smaller developers and open research groups may be excluded before they can compete. The design question is therefore not whether a FINRA-style structure is useful, but whether its governance, access rules and appeal routes keep the rulebook from protecting its authors.