The Daily — 30 September 2026
Trump, Anthropic and who gets to police AI
Trump, Anthropic and who gets to police AI
Trump, Anthropic and who gets to police AI
Donald Trump met some of the most powerful people in artificial intelligence at the White House on Tuesday. Dario Amodei of Anthropic was there alongside Mark Zuckerberg, Sundar Pichai, Nvidia’s Jensen Huang, OpenAI president Greg Brockman, Elon Musk and others. By the end of the meeting the companies had signed a voluntary agreement setting out standards for the safety of advanced AI systems.
The companies agreed to strengthen internal controls and use outside auditors to test whether their systems behave as intended. Their boards would review the findings. Trump described the agreement as something like a constitution for the industry and called it morally binding. House Speaker Mike Johnson was clearer about its legal status: the commitments are voluntary. Congress announced no enforcement mechanism and Trump floated the possibility of a ten-person committee without saying who would sit on it or what powers it would have.
That distinction matters. An outside auditor can examine what a company is doing and report what it finds. A regulator can require a company to meet a standard, demand information, impose penalties and in some cases stop an activity. Tuesday’s agreement strengthens scrutiny but it does not yet create that kind of public authority.
The meeting becomes more interesting when placed beside what Dario Amodei has been saying.
Amodei is not calling for AI development to stop. He argues that development should be paced so safety work can keep up with increasingly capable systems. His proposal would give independent evaluators continuing access inside frontier AI companies, including access to tools, staff and information needed to examine how models are trained and tested. But he goes further than voluntary oversight. He argues that regulation covering all American frontier laboratories would be the most effective approach because it would also bind companies unwilling to cooperate voluntarily.
That puts Anthropic in an unusual position. It is one of the companies building the technology and stands to gain from its spread, but its chief executive is also arguing that companies should not be left entirely to decide their own limits.
Anthropic’s dispute with the Pentagon earlier this year makes the problem more concrete. Anthropic strongly supports the use of AI by the American military and intelligence agencies. Claude is already used for intelligence analysis, operational planning, cyber operations and other national-security work. But the company drew two lines: it opposed using Claude for mass domestic surveillance and said present systems were not reliable enough for fully autonomous weapons. Anthropic also acknowledged something important in its own statement: military decisions belong to the government, not to a private technology company.
That leaves a difficult question. If a company believes a use of its technology is unsafe or incompatible with democratic rights, who should make the final decision—the company, the elected government, Congress, a regulator or a court?
The Trump administration approaches the issue from another direction. Its 2025 AI Action Plan describes artificial intelligence as a race for global dominance. It calls for removing regulatory barriers, speeding private-sector innovation, building data centres and new energy infrastructure and exporting the American AI technology stack abroad. The document also contains work on AI safety and evaluation, but its central aim is clear: the United States should develop and deploy AI faster than its rivals.
So Tuesday’s meeting was not simply a confrontation between an industry demanding freedom and a government demanding restraint. In many respects they want the same thing. The administration wants American AI companies to move quickly. The companies want to build increasingly capable systems. Amodei himself argues that maintaining an American lead matters for national security.
The disagreement is over the rules governing that race.
And that is where the voluntary agreement looks less like a settlement than the beginning of a political argument. The people developing the systems possess much of the technical knowledge needed to understand their risks. Government has powers that private companies do not: it can make law, compel disclosure, set standards and impose sanctions. But government is also actively trying to accelerate the industry and increasingly depends on its systems for military, intelligence and public functions.
The question therefore is not simply whether AI companies can be trusted.
It is how democratic societies build institutions capable of governing a technology when the companies, the state and the national-security system all have strong reasons to make it more powerful.
Tuesday’s agreement provides auditors and promises. It does not yet answer who ultimately has the authority to say no.