Trump's new AI task force splits its leadership among four co-chairs from competing camps

A product lead is about to approve an agent that logs into a customer's systems and takes actions without asking. Her lawyer will want to know what Washington expects. A new White House AI task force looks like the place that answer will come from ^[1].

The hopeful assumption is that one body means one direction. The leadership suggests otherwise. The four co-chairs come from the national intelligence directorate, the Office of Personnel Management, the chief technology officer's post and the FTC, and Fast Company reports that the group brings together officials representing competing forces on AI ^[1][2]. That report links a stated position to only one of them ^[14]: Scott Kupor has argued that detailed rules are a poor fit for a technology moving this fast ^[3]. The report does not describe the task force's mandate or when it will produce anything.

The speed camp has already won some things. The administration has tried to head off tougher state AI laws and has pushed for a "minimally burdensome" federal framework ^[4]. David Sacks, who helped shape the White House AI agenda, argues for building more capable systems faster ^[5].

The pressure for enforcement comes from inside the president's own coalition. In May, more than 60 Trump allies signed a letter coordinated by Humans First. It asked for mandatory testing, vetting and government approval before potentially dangerous frontier models can be deployed ^[6]. Steve Bannon's main worry is job losses. Others in that wing care more about power concentrated in a few large tech companies ^[7]. In the Senate, Josh Hawley and Chris Murphy are drafting a bill that would attach civil and criminal liability to AI agents hacking computer systems ^[8]. Majority Leader John Thune has argued for guardrails ^[9].

China hawks pull both ways. They worry that brakes on U.S. labs would hand Beijing an advantage ^[10]. Treasury Secretary Scott Bessent has also taken an aggressive line on Chinese labs that use distillation to piggyback on U.S. models, and National Cyber Director Sean Cairncross has pushed for tougher action on Chinese AI companies ^[11]. For a product team, those officials help if the product runs on a U.S. model. They become a problem if it depends on a Chinese one.

The model vendors are split as well. According to Fast Company, OpenAI has argued, in effect, that accepting some harms may be part of building more capable AI ^[12]. Anthropic talks about "pacing the frontier" ^[13]. A team's upstream provider may be lobbying for a rule the team would not pick for itself.

So the useful planning unit is the feature, and a 2x2 does the sorting. One axis is whether the feature acts without a person approving each step. The other is whether it depends on a frontier model that an approval rule could hold back, or on a model from a Chinese lab. A feature with a human approving each step, built on a U.S. model below the frontier, is the kind of product the "minimally burdensome" push was aimed at ^[4]. An autonomous agent on a contested model sits under both the Hawley-Murphy draft and the Humans First demand ^[8][6]. I would put compliance engineering into the autonomous column first. The tradeoff is real. Moving a feature toward the safer corner usually means adding an approval step, and users pay for that in slower time-to-value, all to prepare for a liability bill that is still a draft ^[8].