A constitution, in the American tradition, is a restraint on power. The White House has just applied the word to something closer to the opposite.
On Sept. 29, technology executives, including Anthropic’s Dario Amodei, signed a voluntary agreement setting safety baselines for powerful artificial intelligence systems. President Trump called the arrangement a constitution for the industry and said it was morally binding. The same day he announced a federal AI task force, the Super Intelligence Force, to coordinate the preservation of American leadership and the protection of Americans’ interests. Former FTC chair Lina Khan answered that voluntary pledges have failed before, and that Congress should write AI rules while authorities investigate company practices.
The conventional argument will now split into two familiar camps: those who treat a voluntary accord as proof that Washington can stay light-handed, and those who treat every new technology as a jurisdiction the administrative state has not yet fully opened. Both camps are looking at the wrong object.
The object is permission.
A voluntary standard can be a perfectly conservative instrument. Firms should be free to bind themselves, customers should be free to punish recklessness, and the state should not appoint itself editor of every general-purpose tool. But a standard stops being a private discipline when the president declares it morally binding national policy and stands up an executive force to coordinate the field. Nothing in that sequence passed through Congress. Nothing in it can be repealed by the customers who will live with the result. The next administration, of either party, inherits not a press release but a claim: that the safety of advanced AI is an executive portfolio, co-authored with the laboratories large enough to be invited.
Anthropic’s own record makes the co-authorship impossible to ignore. Lobbying disclosures show the company has pressed for mandatory federal oversight. Amodei has called for testing and auditing rules capable of stopping unsafe systems from being deployed. He signed the voluntary accord anyway. Critics, among them David Sacks and Brian Chau, argue that this style of regime would disadvantage smaller competitors. Amodei denies that. The denial is not an argument against the incentive.
Pre-deployment audits with a federal stop switch are not a neutral safety technique. They are a gate. The firms already in the room — staffed, lawyered, and fluent in the baseline they helped describe — can walk through it. A smaller lab with a stronger model and a thinner compliance department cannot, or cannot in time. Competition then shifts from who builds the best system to who is cleared to ship one. That is not free enterprise. It is corporate privilege with a safety letterhead. The privilege does not become public-spirited because the beneficiary denies the advantage.
Khan is not a useful corrective. She is right that self-regulation by the last generation of technology giants did not prevent harms. She is wrong about the moral of that history. Social media’s failures do not establish that Washington should license the next general-purpose technology in advance, or that investigation should become the default posture toward firms whose product is still being invented. Her remedy and the White House ceremony point in the same structural direction: more centralized authority over which systems may exist, with the largest companies either writing the rules or learning to live as their permanent counterparties.
None of this requires pretending the state has no interest. Conservatives who cannot tell limited government from weak government will miss the actual security stake. American leadership in AI is not a slogan. If the systems that touch finance, infrastructure, intelligence, and information are built under rival legal and political regimes, Americans will feel that long before they feel a task force. Energy, talent, serious procurement, and export controls where the national interest is clear are core public functions. A government that shrugs at models that can be turned into fraud engines or weapons is not exercising restraint. It is abdicating.
The error is using that legitimate interest as the door for industrial licensing. A force created to preserve leadership can, without a statute, become the venue where incumbents and officials decide what unsafe means and who is too small to comply. Safety language travels easily. Permission, once granted as a federal power, rarely travels back.
So the debate worth having is not voluntary virtue against mandatory recklessness. It is whether the country will govern concrete harms — crime, fraud, espionage, attacks on critical systems — under ordinary law, or whether it will accept a constitution written at a signing table. The first path holds builders responsible to customers and to courts. The second holds the public responsible to a handshake.
If Congress writes rules, it should write rules aimed at harms it can define, not a general authority to halt deployment because a baseline, drafted with the regulated, says stop. If the executive wants American advantage, it should compete. It should not baptize an industry accord as higher law.
The part conventional coverage will understate is the precedent hiding inside the courtesy. Voluntary was presented as the alternative to federal oversight. Watched closely, it was the announcement that federal oversight had found its authors — and that the authors include the firms with the most to gain from a gate.
