The Self-Preflight

When a regulator's first draft is written by the firm it is meant to discipline, the test is usually called a partnership. Call this the self-preflight: the regulated entity designs the test, runs the trial period, and writes the conditions under which its own compliance becomes mandatory. The concessions that look like safety wins are the ratchet that locks the regime in.

Demis Hassabis just endorsed a preflight safety testing regime for frontier AI. Frontier labs would share new models with a Standards Body up to thirty days before release; once the body's assessment protocol is shown to be effective and robust, formalisation could quickly follow, and frontier models would be required to pass it to ship in the US market. The path to mandatoriness runs through voluntary self-rule.

FINRA, the broker-dealer self-regulator Hassabis and Gary Marcus both invoke, was supposed to police Wall Street in the public interest. It ended up writing the exams the firms wanted and absorbing the cost of its own capture. Pharma's post-market surveillance, platform trust-and-safety, and the early nuclear safety committees tell the same story: a regime born of good faith becomes the lab's moat, because the firms that can afford to staff the standards body set the floor everyone else must clear.

The labs gain the rulebook. The public gains whatever the labs decided the public was going to get.