Supreme Court ruling in *Trump v. Slaughter* threatens independent AI oversight by putting federal agencies under direct presidential control

Imagine Congress creates a Frontier AI Commission that can license major training runs, compel evaluations, restrict deployments, order emergency pauses, or impose penalties. Under today’s decision, its leaders could not be protected from presidential removal.

@benrossen Very bold to assume these "independent" bodies won't be captured by partisan bureaucrats in your counter-factual world.

@bradrcarson My (personal) opinion is that federal licensing authority is probably a good thing to avoid the kind of situation we’re in now, but I think it needs to be done in a way preserves some semblance of independence.

@benrossen >Imagine Congress creates a Frontier AI Commission

See, that right there is where you went wrong

Congress shouldn't do that

The Federal Reserve is not a promising template. It prevailed because of a specific tradition predating the Constitution and a unique historical lineage through the First and Second Banks. Frontier AI has no comparable pedigree.

Accountability measures through private auditors, subject to standards and accreditation established by Congress or via rulemaking, and reporting up to CAISI, wouldd also be sensible.

The U.S. already has the institutional core of the technical side: the US CAISI. It is intended to develop AI measurement science, evaluate advanced models, work with developers, coordinate across government, and shape standards.

Why? The Court treated substantive rulemaking, investigations, enforcement, civil litigation, and in-house adjudication as executive power. A principal officer wielding those powers must be removable by the President at will.

Of course, CAISI is not constitutionally “independent.” It sits inside Commerce. But that's not the point. Its role should be fundamentally scientific, not coercive: tell us what models can do, how much confidence to place in the evidence, and how risks should be measured.

This would not remove politics from AI enforcement. Slaughter largely forecloses that. But it could go a long way towards a shared technical baseline that is harder to bend toward a preferred policy outcome.

But: Slaughter does not make sound institutional design impossible. It points toward a split architecture: separate technical measurement and assessment from the government’s coercive licensing and enforcement powers.

A general understanding that independence might be helpful is not itself enough: otherwise the same reasoning would protect the FTC, CFPB, NLRB, etc. The majority squarely rejected this technocratic argument.

In a split design, CAISI could run or validate evaluations, build benchmarks, develop measurement standards, publish findings, and maintain secure channels for model access. Its work would inform, but not itself dictate, licensing or enforcement.

@benrossen I read the Constitution. I don't see a fourth branch of government in there.

@benrossen There are 2 types of Federal bodies only: the political & those that are fundamentally dishonest about not being political.

If it wasn’t political the Democrats wouldn’t be apoplectic & Trump wouldn’t care about removing her.

See also all “non-partisan” bodies & roles. Lies.

@benrossen yes I'm sure that "impartial commission" will not be heavily captured from the start

@benrossen Helpful thoughts. Is your opposition then to a federal agency with licensing authority or is that you want the technical standards-setting to be separated from this agency?

@benrossen >Imagine Congress creates a Frontier AI Commission

You somehow think that some bureaucrats should be immune from the will of the people

@benrossen "free from political or partisan influence"??? what planet are you living on?

@benrossen “federal body free from political or partisan influence.” 👈 should never have been a thing, has no basis in the Constitution, and is sheer insanity. No unelected bureaucrat should exist as a fourth column of government. Fucking shitbags who think this was EVER okay.