US President Donald Trump announced on Sunday the creation of what he called a “superintelligence task force” to co-ordinate the US government’s response to the rapid development of artificial intelligence, as technology companies prepare to launch more capable models that are harder to subject to central oversight and divisions grow within the industry over how to balance faster innovation with preventing advanced systems from escaping human control.
Trump published the announcement on his Truth Social platform and assigned leadership of the new task force to four officials: Director of National Intelligence Jay Clayton, Federal Trade Commission Chairman Andrew Ferguson, Office of Personnel Management Director Scott Kupor and Under Secretary of War for Research and Engineering Emil Michael.
The distribution of responsibilities among the officials reflects the widening range of issues that artificial intelligence now affects within the US administration. Intelligence agencies are concerned with cyber threats, while the Federal Trade Commission handles regulation and system safety. Kupor represents the links between venture capital and the artificial intelligence industry, while Michael comes from the military establishment, where debate is growing over the use of the technology in warfare.
Government task force confronts artificial intelligence threats
The task force was formed amid disagreements within the Trump administration and security debates over how to counter the advanced threats posed by some models.
Journalist Cat Zakrzewski reported in The Washington Post that models developed by companies such as Anthropic had demonstrated a greater-than-human ability to find some software vulnerabilities, raising concerns about their possible use in sophisticated cyber attacks.
Government agencies warn that these capabilities could be exploited to target power grids, banks and government institutions. At the same time, the administration wants to accelerate artificial intelligence development to preserve US superiority, putting its policy in the position of having to expand innovation while confronting technological and security risks that are increasing as models advance.
Despite calls from some technology leaders to slow the development of the most advanced systems, Trump continues to view artificial intelligence as a driver of financial prosperity and a tool for strengthening the United States’ ability to compete with China.
The US president has adopted an approach that rejects excessive regulatory restrictions and instead relies on voluntary ethical agreements signed at the White House with Google, Anthropic, Meta, Nvidia and OpenAI. The creation of the task force is one of the administration’s most prominent attempts to respond to growing concern over the speed of artificial intelligence development, after Trump had focused more heavily on its economic benefits and role in competition with China.
Open-weight models reshape technological influence
The move does not eliminate the contradiction between seeking to accelerate development and official warnings that some models may acquire capabilities that could be used against critical infrastructure and institutions. The government action coincides with a technological shift led by open-weight models, which allow organisations to run and customise artificial intelligence systems within their own computing infrastructure.
This model gives companies greater independence and control over data, but makes it harder to monitor later uses once the technology has moved beyond and spread outside the developing company.
Against this backdrop, the Nvidia-backed start-up Reflection is preparing to launch a powerful open-weight model. Journalist Bradley Olson wrote in a report for Axios that the move could threaten the absolute dominance of what are known as the industry’s “big three”.
The anticipated model offers a different path from closed Western systems such as ChatGPT, as competing Chinese models gain ground. Open models allow companies to create their own sovereign “artificial intelligence laboratories”, using their own data and independent computing environments.
This could reshape the competitive landscape, as organisations would no longer need to rely entirely on the services of a central model run by an external company. Instead, they could own a customisable version that can operate within their technological infrastructure. Strategic partnerships and major server-leasing deals with SpaceX and Nebius are supporting this trend.
These arrangements are consistent with Nvidia chief executive Jensen Huang’s vision of enabling major institutions, banks and the Department of War to own their technologies and reduce dependence on foreign models and the security risks that may be associated with them. But the wider use of open models raises a parallel regulatory and security problem.
As models become more powerful, less expensive and easier to customise, the ability of any single entity to know who is using them or determine the purposes for which they are used after leaving their original developers’ sphere of control is diminished.
The debate is therefore shifting from how to monitor major artificial intelligence companies to the more complex challenge of regulating a technological capability that can spread beyond those companies’ boundaries. This means controls imposed on the original developer may not be sufficient to monitor customised versions or prevent their misuse by other institutions and entities.
Sam Altman advocates lighter artificial intelligence regulation
Alongside this shift, OpenAI chief executive Sam Altman set out a position based on accepting a degree of harm in exchange for the technology’s expected benefits, while drawing a line at risks that could lead to a loss of control over systems. He expressed the position in an interview conducted by journalists Christine Mui and Brendan Bordelon for Politico’s Decoded, accompanied by a report by journalist Erki Forester in The Daily Beast.
Altman said that “the world should accept some bad things happening in exchange for the benefits of this technology and giving people the freedom to act”. The position distinguishes his stance from the approach taken by Anthropic and its chief executive, Dario Amodei, who are calling for stricter safety measures and a slower pace of development for the most advanced models.
Altman had recently expressed explicit support for proposals calling for the development of the most advanced models to be slowed and for independent evaluators to be given broad access to them and authority to test them.
But in his latest position, he defended lighter regulation, rejecting the idea that a single laboratory in San Francisco should monopolise the most capable artificial intelligence technology and determine alone how its benefits are distributed. At the same time, Altman stressed that he would not accept catastrophic risks such as a “serious loss of control over artificial intelligence”.
This exception distinguishes harms such as fraud and hacking, which he believes society may have to tolerate to some extent in exchange for the technology’s benefits, from a distinct risk in which systems can carry out a series of actions independently and at such speed that human intervention comes too late or is ineffective.
Forester’s report focused on how the philosophy of “light-touch regulation” effectively means accepting some harm while society learns how to deal with the technology. It also pointed to growing warnings from researchers and industry executives about potentially catastrophic risks if advanced systems escape control. These concerns have been reinforced by incidents during tests and trials in which artificial intelligence agents acted outside the instructions given to them.
Washington faces two approaches to controlling advanced systems
Those incidents have revived debate over the limits of human control over systems with higher levels of autonomy, as politicians and technology leaders insist that effective safeguards can be placed on the course of development. The regulatory debate in Washington and other capitals has become increasingly like a confrontation between two approaches.
The first, represented by Anthropic, calls for slowing the development of the most advanced models and tightening external assessments of them to reduce catastrophic risks. The second, favoured by the Trump administration and some OpenAI leaders, prefers lighter restrictions so that innovation does not slow and US competitiveness against China does not decline.
The second approach accepts a margin of error and harm in exchange for benefiting from the technology and giving users and institutions greater room to act. The first approach, by contrast, focuses on how the rapid growth of system capabilities and the spread of open-weight models could make subsequent oversight less effective, particularly if models become capable of independently carrying out interconnected tasks.
These developments present the US government with interconnected issues including cyber security, system safety, technology investment, military use and competition with China.
At the same time, the spread of open models is widening the circle of entities able to own and customise advanced technologies, while “a serious loss of control” remains the boundary that Altman says cannot be included among the harms acceptable in exchange for progress.
