The officials and the executives had been meeting for weeks in Washington, and by Thursday an outline had finally taken shape: the government would not impose binding rules, and the companies would accept a set of voluntary standards that could be published as soon as next week, according to people familiar with the discussions. The framework, still being negotiated, is the administration’s most concrete attempt yet to give the frontier-AI industry something it has been asking for — predictability.
The details under discussion are specific. Government officials and the leading AI companies are debating when a model should be classified as “frontier,” how long safety evaluations should last, and what safety standards a model must meet before it is deployed publicly, according to people familiar with the matter. A fourth question has proven the most delicate: who should be allowed to access the most advanced models, in the United States and abroad. That last item, officials say, could form the basis for a broader framework among U.S. allies.
The voluntary approach reflects the administration’s preference for light-touch regulation. Rather than legislation or binding agency rules, the White House has pushed for a set of commitments that companies adopt on their own, with the government defining what good practice looks like. Industry executives have largely welcomed the direction, arguing that voluntary standards give them flexibility while still addressing the safety concerns that have driven calls for regulation. The alternative, they warn, is a patchwork of state laws and agency actions that would be harder to comply with.
The timing is not coincidental. The companies sitting at the table have all faced moments in recent months when the absence of clear rules created chaos. One leading lab shut down access to its most capable model after a vulnerability was exploited, then restored it weeks later after negotiating new safeguards with the government. Another was reported to have been asked by the White House to delay a product release. In both cases, the industry’s complaint was the same: the rules were being made up as events unfolded, and no one knew where the line was.
The access question is where the discussions get hardest. The most advanced AI models are already treated as strategic assets, with the government weighing export controls and the industry weighing international customers. A voluntary standard that says who may use the most powerful systems — which countries, which companies, which researchers — would go further than any existing agreement, and people familiar with the talks said the language is still being drafted. The ambition is a shared framework with allied governments, so that a model approved in Washington is accepted in Tokyo and London.
Skeptics have their doubts. Voluntary standards, critics note, are only as strong as the willingness to enforce them, and the companies that sign up are the ones that get to decide whether they have complied. Previous voluntary commitments — the safety pledges that major labs made in earlier years, and the executive order on AI safety issued by the prior administration — produced frameworks that were never tested under pressure. The new effort, they argue, risks repeating the pattern: a document that looks serious and binds no one.
Supporters counter that the alternative is worse. AI is developing too fast for legislation, which takes years, and too fast for agency rulemaking, which takes longer. A voluntary standard published next week can be updated next month, and it gives companies a public benchmark against which their behavior can be measured. The administration also has a lever that earlier efforts lacked: the government is a customer of AI services, a gatekeeper for exports, and a partner in infrastructure projects, and all of those roles can be used to encourage compliance without legislation.
For the companies, the stakes are commercial as much as political. A published framework would give them cover for decisions they already make — when to release a model, how long to test it, whom to serve. It would also give them a defense against the most aggressive state proposals, which have included liability rules and mandatory audits. Industry lawyers have told clients that a credible federal framework, even voluntary, would make it harder for states to impose their own requirements.
People familiar with the talks said the remaining disagreements are narrow but real: the definition of frontier, the length of evaluations, and the enforcement of the access rules. All three are expected to be resolved in the coming days, with an announcement possible as early as next week. The companies involved declined to comment on the specifics, citing the sensitivity of the negotiations.
What emerges will be watched closely beyond Washington. European regulators, who have pursued binding rules through the EU AI Act, are following the American experiment with interest, and allied governments have told U.S. officials they want to be consulted before anything is published. If the voluntary framework works — if companies adopt it, if the definitions hold, if the access rules survive contact with the market — it could become the template for AI governance across the democratic world. If it fails, the vacuum it leaves will be filled by someone else.
