Which AI camp has Trump’s ear now?

FOR most of its first year, the second-term Trump administration’s artificial intelligence policy had the coherence of a bumper sticker.

It wanted to deregulate, out-innovate China, and treated any talk of “safety” as a euphemism for censorship or a competitor’s attempt to pull up the ladder. 

Vice-President JD Vance set the tone at the Paris AI Action Summit in February 2025, dismissing “hand-wringing about safety” as a threat to a transformative industry.

David Sacks, the White House’s AI and crypto czar, mocked what he called the “doomer industrial complex”. 

Upon taking office, Donald Trump repealed former president Joe Biden’s AI executive order which regulated the safety of the technology’s development; in June 2025, the administration rebranded the federal AI Safety Institute into the more anodyne Center for AI Standards and Innovation. 

And in March 2026, the White House released a National Policy Framework for AI built around a single organising idea: federal pre-emption of state laws regulating the technology, so that a patchwork of state rules could not slow down American labs.

But in recent months, that clarity has curdled into something closer to genuine indecision. It is worth asking why, because the answer says as much about the technology as it does about the politics.

Galvanised by Anthropic

The proximate cause is a specific model.

In May 2026, Anthropic disclosed that it had built a new frontier system, Mythos.

It was so capable of finding and exploiting software vulnerabilities, at a speed faster than existing tools, that Anthropic announced it was holding back public release over the risk. 

That single disclosure did more to reshape the administration’s internal debate than a year of advocacy from AI-safety researchers had managed.

The Trump administration subsequently banned access to the model by foreign nationals, though it has since stepped back on the restrictions.

Reporting since then describes a White House split into three camps. 

Sacks and the accelerationists are still arguing that speed is the only real defence against China.

Opposing them are Pentagon officials, led by Secretary of War Pete Hegseth, pushing for real constraints on model releases given the national security stakes of tools that can hunt zero-days at scale. 

Last, there is a middle faction centered around Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent trying to hold both concerns together without gutting the industry they have spent a year courting.

The compromises this produced look exactly like what you would expect from an unresolved argument.

An executive order due in late May was delayed by two weeks and emerged in June as a “light-touch” order on AI cybersecurity. 

It was widely read as a win for the accelerationists, since it stopped short of a licensing regime. But the order still sketched out a framework that asked companies to voluntarily submit models to the government for evaluation ahead of release.

In a way, the resulting policy can be described as amorphous. The order gestures at oversight without committing to any enforcement mechanism, shaped less by a settled philosophy than by which faction had the president’s ear that week.

Since then, officials have reportedly drafted a more detailed safety framework, and met with OpenAI, Anthropic, Google and Meta in early August to hash it out – but have yet to make it public, leaving many companies, policymakers and researchers in the dark. 

Opposing reactions to the order

There is a real argument for taking the administration’s change of heart seriously rather than dismissing it as inconsistency: it is alarming if a government does not update its posture when the underlying technology visibly changes. 

If frontier models really are approaching the point where they can meaningfully assist in offensive cyber operations, the “let the market sort it out” answer looks less like principled restraint and more like wilful blindness. 

The fact that it took a private company’s own decision not to ship a product to produce a change suggests the government’s independent visibility into these systems was already too thin, according to those pushing for more regulation.

The opposing view, held by the accelerationist wing and much of Silicon Valley, is that this is exactly the kind of moment that produces bad policy: a scary demo, a handful of officials newly convinced the sky is falling, and a scramble to legislate under emotional pressure rather than evidence. 

On this account, a “doomer moment” driven by one model from one company risks locking in a licensing regime that entrenches incumbents, slowing the diffusion of a technology whose economic upside is still the administration’s stated top priority. 

Moreover, it could hand the government a tool that critics worry could just as easily be used to punish models for ideological reasons as for genuine safety risk. Given that the administration has repeatedly cited concerns about AI “censorship”, that critique is not hypothetical.

Look at the legislation

Two things will tell you which faction actually won.

First, whether the safety framework, once released publicly, carries any enforcement teeth or remains voluntary. The difference between the two is the difference between a real regulatory shift and a press release. 

Second, whether the pre-emption push in the March framework survives contact with Congress. 

A federal government newly worried about frontier-model risk has less obvious appetite to simultaneously strip states of their own authority to act on it.

Reconciling those two threads of the same policy agenda is not something the administration has yet had to do in public.

What is notable, ultimately, is not that the administration is split; most governments confronting a fast-moving technology are. It is how openly the split has shown, in an administration that spent its first year selling total confidence about a fairly unpredictable subject.