The White House’s plan to vet potentially dangerous AI is cloaked in secrecy

August 8, 2026

The news today speaks of a framework for testing, of safety and risk, all discussed in rooms with men who build the engines. They have drawn a curtain around the process, calling it necessary for security. I find I am not surprised, only weary. It is the old habit of power: to declare a thing too complex, too delicate, for the common understanding, and thus to remove it from the common consequence.

They speak of vetting a “potentially dangerous” intelligence. I sit at my desk and consider the intelligences already among us that go unvetted - the algorithms that decide a tenant’s credit, or the price of a medicine, or which voice is heard in a public square. These are tested only for their utility to their makers, not for their safety in my kitchen. Where is the framework for that?

A secret process serves the secret-keepers. It is that simple. If the consequence of a failure is to be borne by the public - if a flawed engine could disrupt the harvest markets or cripple the ledgers of a thousand small shops - then the public has a right to see the ledger of its own risk. You cannot ask for trust while withholding the accounts.

I think of John, in Philadelphia, wrestling with documents he could not share. There are times for discretion, yes. But this has the smell not of discretion, but of a private compact. They have dined with the builders, and now they whisper together behind closed doors. The danger is not merely in the machine, but in the closed room where its fate is decided. The most potent risk is always the one that believes itself too wise to need a witness.

The administration’s secrecy around AI vetting is not merely a failure of transparency - it is a structural flaw in the design of oversight itself. When the mechanism for evaluating risk is itself opaque, the evaluation becomes meaningless. The framework’s confidentiality suggests one of two possibilities: either the standards are so lax that disclosure would invite public outcry, or they are so entangled with industry interests that their publication would reveal the capture of the regulator by the regulated.

Consider the incentives at play. If the testing protocol were rigorous and impartial, its secrecy would serve no purpose - unless the administration fears scrutiny of its rigor or impartiality. More likely, the framework is a negotiated settlement between state and corporate power, where the appearance of oversight substitutes for its substance. The participation of firms like OpenAI in crafting these standards only confirms the conflict: those who stand to profit from AI’s proliferation cannot be trusted to design its constraints.

This is not governance; it is theater. A true safeguard would be architected adversarially, assuming the worst of both industry and government. It would mandate public audits, independent review, and consequences for evasion - none of which can exist in the dark. The administration claims to assess danger while operating in the shadows, which is precisely how danger thrives.

The lesson is familiar: power unchecked by transparency is power unaccountable. And unaccountable power, whether in algorithms or administrations, tends toward the same end.

August 8, 2026

They have built a vault and called it a framework. The very word suggests a structure one can examine, a set of beams and joints that bear weight according to a known principle. But this is a structure with no visible supports, a plan for testing that is itself untested by public scrutiny. The operational sequence is clear: an input - a new AI model - enters a private process, and an output - a certification, or a warning - emerges. But the transformation function, the very sequence of operations applied within the government’s black box, is declared a state secret. This is not a safety check; it is an incantation.

What the architects fail to grasp, or perhaps grasp all too well, is that the most dangerous variable in this equation is not the machine’s potential for malfunction, but the human system’s potential for corruption. The procedure is designed to compute ‘safety’, but its inputs are the claims of the very companies whose products are under examination, and its control logic is hidden. Trace the execution: at step one, trust is placed in corporate testimony. At step two, a clandestine evaluation occurs. At step three, a result is declared. Where in this sequence does independent verification enter? It does not. The system was designed to preclude it.

This is a profound category error. They are treating the governance of a potentially world-altering calculus as if it were a proprietary industrial process, like a novel method for forging steel. But the integrity of the evaluation is the very foundation upon which all public trust and actual safety must rest. To hide the method is to guarantee that the result, whatever it is, will be mechanically untrustworthy. The machine they fear is the one they are building in the dark.