A couple of professors at the University of Pennsylvania’s Wharton School Department of Legal Studies and Business Ethics want a body of Experts to supervise risks from emergent AI, saying that such a body would be better than an FDA-like regulatory body, or Congress through statutorily enabled product-safety laws.

They’re right that having a government body of experts like the FDA do this sort of thing is determinedly suboptimal. They’re right, also, regarding Congress, although Congress is considerably more malleable than a department or agency of bureaucrats.

But another body of Experts?

Bank supervision, which emerged in the Civil War and took its current form out of the Great Depression, offers the best framework for overseeing the most advanced AI labs.

After all,

Frontier AI labs such as OpenAI, Anthropic, and Google DeepMind are different.

And there’s always an excuse for standing up yet another bureaucratic regulatory body. In the case of their bank supervision model, about which they’re so enthusiastic, they give their game away [emphasis added].

Banks are too complex to govern through legislated rules alone, too important to leave to market discipline, and too dynamic for one-time approval.
Bank examiners often sit inside the institutions they oversee.

That’s the problem with our economic system government overlords. With the Panic of 2008, the Federal government created out of whole cloth the myth of some (ultimately government favored) businesses are too big to fail and so must always be guaranteed a government bailout. That confidence in the Federal apparatchiks sitting inside the banks also is misplaced. It’s only necessary to see the failures of the Silvergate, Signature, and First Republic banks to see the intrinsic failure of this. Those banks didn’t only fail through their own mismanagement; they also failed because their regulators were incompetent enough or lazy enough or complacent enough to miss those bank managers’ basic economics error of borrowing short-term while lending long-term and letting those two get ‘way out of balance. That allowed their short-term debts to come due before they had the long-term debt income to cover.

But the good professors want a board of Expert Apparatchiks inside the OpenAIs, Anthropics, and DeepMinds to oversee how these handle risks of emergent AI.

And this:

Banks share information with supervisors that they could never safely disclose publicly.

AI software is too important, too critical to national security, to share with apparatchiks of government. Our Federal government is infamous for its inability to defend against PRC cyber espionage. It’s infamous, also, for its bureaucrat employees leaking confidential financial data about businesses and persons of which those bureaucrats personally disapprove.

And this:

An AI risk supervisor could be funded by industry fees. Its leaders should be Senate-confirmed and removable by the president, but its expert staff should be insulated from day-to-day political pressure.

No.

Experts have their uses, often very important uses. On the witness stand to explain this or that aspect of a crime, balanced by another expert on the witness stand with a differing explanation. In police department forensics sections. As teachers in environments where their expertise is more important to the teaching than their teaching style. In medical and mental health doctor offices.

But in government? Not so much. Experts are useful when they’re part of a range of experts advising, as employees, the government’s decision makers. But as government decision makers? Definitely not so much. For the lack of utility of that last role, it’s only necessary to look at the Fauci-Collins-led experts as bureaucrats, or at the experts of the John Brennan and James Clapper CIA and ODNI, respectively.

The Supreme Court was right when it greatly reduced Chevon Deference in its Loper Bright Enterprises v. Raimondo ruling, making clear that “experts” in government aren’t owed any particular deference on matters of government behaviors and decision-making.

So it is with emergent AI.