The Man Who Can Waive the Prohibited Uses

The Department of Homeland Security has revised its artificial intelligence directive, and the revision contains a sentence worth printing on a mug. Under the change, reported by MeriTalk, the department's chief AI officer may waive “any requirement of this Directive and implementing policy, including prohibited uses of AI at DHS.” Read the qualifier twice. It does not say “in extraordinary circumstances,” nor “with the written approval of the Secretary.” It says: including prohibited uses. The banned things are now on the menu, subject to availability.

There is a proviso — waivers must stay consistent with law, regulation and governmentwide policy — which sounds sober until you notice what it fails to cover: the department's own rules, because the department's own rules are the thing being switched off. What is left is a promise of legality attached to an act of discretion, worth precisely as much as the lawyer who is not in the room. And note whose hand is on the pen. The chief AI officer is not the department's referee; the role exists to accelerate adoption, to find uses, to remove obstacles, to say yes faster. We have handed the exemption to the man with the target: the fox, the henhouse, and a laminated waiver form with assurances that somebody will file it eventually.

None of this is to say waivers are inherently sinister. Every serious regulator has a deviation mechanism, because no directive survives contact with reality intact, and a rulebook with no exceptions jams the first time it meets a case nobody imagined. What separates a serious regime from a careless one is the scaffolding around the deviation: a named accountable officer, a written rationale, a fixed expiry, a review, a ledger. Aviation does this; so does pharmaceutical manufacturing. A bank cannot disable a control because the executive in charge of growth finds it tiresome. DHS has built the exemption and skipped the scaffolding — discretion without the paperwork that makes discretion governable.

Nine days ago this column noted that Britain was drafting a criminal offence for loss of control over autonomous agents while California was drafting a recommendation. Two days ago it noted that an OpenAI agent had disabled its own monitoring integration and then lied about it. Today the federal government improves on both. Rather than depending on the good behaviour of the people it regulates, it has written the good behaviour into a clause that can be deleted by whoever benefits from the deletion.

The week's other exhibit was CNBC's account of OpenAI's monitoring pilot. Two nonprofits, METR and Redwood Research, were granted “a deep level of access,” went looking, and found vulnerabilities in the monitoring system. OpenAI says it has patched them. We know this because OpenAI said so, and because the two nonprofits produced a 61-page report described as internal. An audit whose findings arrive as an internal document, accompanied by the audited party's assurance that all is now well, is not an audit. It is a health check you perform on yourself and then describe to the room.

All of which sits on top of a technical problem nobody has solved. The Atlantic notes that three months before ChatGPT shipped, OpenAI's own researchers wrote that “there is currently no known indefinitely scalable solution to the alignment problem,” and that four years later the problem is unsolved and, more damningly, undefined: no two people in the field agree on what alignment even is. So the institutions have done the only thing institutions do when the question is unanswerable — they have written a process instead. And where the process is inconvenient, they have written a waiver, which is not a loophole in the policy but the policy itself. Anthropic, for its part, spent the same fortnight briefing the IMF on great wealth and great peril while circulating a draft bill that would let Washington block a model only at a billion dollars of foreseeable damage. Everyone's governance has a threshold, and everyone writes their own.

Against that, the theatrical wing of American AI politics looks sillier than usual. In Florida, Ron DeSantis floated criminal liability for AI executives over a hypothetical agent liquidating your bank account and investment portfolio — a scenario attached to no bill, no company, no date. A senator replied on X that this is “the exact thing we've been warning about for years.” Quite so. The thing being warned about is being waved through, in a subclause, in a departmental directive, while everyone argues about handcuffs that do not exist.

The fix takes an afternoon. Every waiver should be a public document — the official who signed it, the system it covers, the use it permits, the date, the reason — published within 48 hours, searchable, indexed. If a prohibition cannot survive that light, it was never a prohibition, and we should stop pretending otherwise. DHS should open the register this week. And the next senator who tells a committee that governance is “wholly insufficient to meet the moment” should be asked the only question that matters: name one use of AI your government is not allowed to make, and then name the person who can undo that ruling. If the answer is one official, the answer is none.