America's leading AI firms are recruiting the state as a competitor

A FEW SHORT MONTHS ago, the Trump administration was still telling investors that artificial intelligence should be left alone. Now it is inviting staffers from major AI companies to the White House to discuss how the government might review their models before they are released. Staffers from OpenAI, Google and Anthropic are expected in Washington on Tuesday to review a framework that was supposed to be voluntary but now looks less so. The real story is not that the White House is changing its tune. It is that the companies with the most to lose from competition are helping write the rules.

On June 2nd, Mr Trump signed an executive order directing federal agencies to design a voluntary framework under which developers of frontier AI models could provide the government with up to 30 days' access before public release. The order explicitly disavowed any licensing or preclearance regime. Nothing, it said, may be construed as authorising mandatory government approval for the development, publication or distribution of AI models. The order arrived after the president had repeatedly promised a hands-off approach, describing the industry as "a beautiful baby that's born" that must be allowed to thrive.

The pivot was not entirely unexpected. The same executive order tasked agencies with benchmarking the cyber capabilities of frontier models and creating a clearinghouse, called Gold Eagle, to coordinate vulnerability detection between government and industry. Cybersecurity is a legitimate national-security concern. Frontier models with advanced reasoning abilities can, in theory, be turned into tools for sophisticated cyberattacks. The administration launched Gold Eagle on July 14th, in partnership with the Treasury Department, the Department of Homeland Security and the Department of War.

Yet what has happened in practice suggests something other than a narrow security review. According to CNBC, citing people familiar with the matter, the administration is now dictating which companies and entities are allowed access to the latest frontier models. Until recently, that decision rested with the companies themselves: Anthropic managed access through its Project Glasswing, and OpenAI through a similar initiative called Daybreak. Going forward, rollouts will reportedly require explicit government approval for which partners can participate. In June, the White House briefly blocked Anthropic's Claude Mythos 5 and Fable 5 models on national-security grounds, reinstating access only after weeks of negotiation. OpenAI said in June it would limit new models to "trusted partners" to comply with government requests.

To be sure, the White House denies any formal control. A White House official told CNBC that engagements remain "voluntary" and that "decisions on timing and scope of releases rest entirely with the companies." The administration's stated aim is to strengthen security without stifling innovation. The concern about cyber risks is not frivolous.

But the deeper problem is who benefits from this arrangement. According to The Information, OpenAI and Anthropic have pushed to expand the type of models that would require government review under the framework to include those from their competitors. The incentive is clear. A regulatory framework, however voluntary in name, raises the cost of entry for smaller labs and foreign rivals. It entrenches the incumbents who have the most access to Washington and the most to gain from being the government's "trusted partners." This is not innovation policy. It is rent-seeking dressed in the language of national security.

The trouble is that the competitive landscape is shifting faster than the bureaucracy can adapt. On July 16th, China's Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that largely caught up to the performance of Fable 5 and GPT-5.6, the latest offerings from Anthropic and OpenAI, and outperformed them on at least one independent benchmark. David Sacks, the former White House AI czar, called the Kimi K3 breakthrough "concerning". "This is how you lose the AI race," he wrote. "The rest of the world won't play by our rules if we bog ourselves down."

That is the paradox the administration faces. A regulatory framework negotiated with a handful of American incumbents may slow them down while doing nothing to stop Chinese labs from advancing. The Kimi K3 release came about two weeks after Claude Fable 5 became available worldwide again on July 1 following a suspension linked to U.S. export restrictions, suggesting that even the attempt to gate models can backfire by disrupting distribution. The framework's deadlines were set for August 1st, 60 days after the executive order. The institutions that would carry it out - the NSA, CISA, and the Treasury Department - were asked to build a classified benchmarking process in that time. One suspects the machinery will be more political than technical.

OpenAI's Sam Altman, who was in Washington in late July to meet with Susie Wiles, the White House chief of staff, and a range of senior officials, has been an enthusiastic participant. But participation carries its own risk. Companies that invite the government into their release process become dependent on it. They also become complicit in a system that their competitors may later exploit. If OpenAI and Anthropic are helping the White House build a framework that raises barriers to entry, they may eventually find that the government they are courting does not distinguish between their rivals and their own products.

The better answer is not a private regulatory club run by the government's most favoured firms. It is to compete faster. The Luddites were wrong about machines in the long run, but right about the pain of transition. The aim should be to lower barriers for American researchers, speed up permitting for data centres, simplify export controls that complicate distribution, and invest in the energy and compute infrastructure that actually determines who wins an AI race. Those are hard things. Regulatory frameworks that benefit the insiders are easy. That is precisely why they should be resisted. The first rule of industrial policy is not to subsidise your favourites. It is to make sure they cannot afford to be favourites for long.