At Jackson Hole, a Princeton economist asked bank regulators to vet frontier models before release

Brunnermeier's remedy points upstream of every entity a bank examiner licenses, because evaluating a major model before it reaches the public ^[3] means standing inside a vendor's release calendar rather than inside the supervised bank, and prudential leverage normally runs the other way, through the charter.

Which is why the enforcement objection is the more interesting half. Sultan Meghji, the CSIS senior technical expert who was previously chief innovation officer at the FDIC, told American Banker that seemingly independent models built or trained in similar ways could produce manipulation without communication or explicit intent, leaving "nothing to subpoena, no data to track and no one to charge with a crime" ^[8]. Brunnermeier's companion point runs alongside it: the central bank is legible to market AI, since its speeches, minutes and entire history are training data, while the AI ecology is not ^[9]. Put those together and the supervisory response most likely to survive contact with an exam cycle is the dull one, which is asking each bank which providers it cannot operate without, and who else depends on the same ones.

Now calibrate the thinness. American Banker headlined five dystopian futures and the text in front of me names four of them, being colluding trading agents, an outsmarted central bank, a kill switch that turns out to be an illusion, and provider concentration, which leaves one scenario I cannot see ^[14]. The room was roughly 120 central bankers, economists and academics ^[4], and the speaker was a professor, not a supervisor holding a consultation paper.

How this reads depends on which path holds. It stays a conference paper, and the visible constituency for follow-through is people like Ken Tower of Quantitative Analysis Service, who wrote on LinkedIn that he hoped the concerns and recommendations would be taken seriously by exchanges and regulators ^[12], and note that he put exchanges first. Or the language lands bolted onto existing third-party risk expectations, which costs a bank a list and a paragraph. Or it lands as resilience testing against the kill-switch scenario, which costs considerably more, because you cannot table-top a market in which algorithmic market makers stop supplying the bulk of liquidity and human traders are expected to replace them ^[10].

My view, and this is probably wrong, is the middle one, on the reasoning that supervisors reach for the tool already in their hand rather than the one that requires standing over a vendor they do not license.

The cheap version is a page: providers, models, what breaks when they go dark, which of your competitors buys from the same firm. The expensive version is routing that page through full model validation. A one-page register costs a week of somebody's time; validation staff spent on procurement paperwork are not spent on the models already in production.