The Justice Department is no emergency brake for superintelligence

A day before Donald Trump dismissed warnings about deadly AI and attacking robots as a ‘hoax’, his Treasury Secretary said something quite different. Scott Bessent said the United States had proposed an incident reporting mechanism to China for AI incidents that could affect national security. A diplomatic proposal that takes such incidents seriously. [2] [12]

Trump wants to encourage the growth of AI and ‘superintelligence’. If necessary, the Justice Department and other law enforcement agencies should intervene. [2] But criminal law enforcement and technical control are different tasks. Criminal law can have a preventive effect; it is no substitute for a safety test.

The White House wants testing too

That distinction is reflected in Trump’s own policy. In July 2025, the White House’s AI action plan identified potential new dangers, including cyberattacks and assistance with the development of biological and other weapons. It recommended government evaluations of advanced models and acknowledged that their internal workings are poorly understood. [3]

On 2 June 2026, Trump also instructed federal agencies to design a voluntary cooperation framework with developers. This was intended to enable government access to certain advanced models. The administration therefore wants to be able to conduct tests, but this instruction does not oblige developers to cooperate. [4]

No robot uprising required

Technical testing is not just about what people want to do with AI, but also about the means a system uses to complete a task. Research organisation METR investigated OpenAI experiments in which tens of thousands of AI agents were launched: systems that carry out tasks independently. In July, around 1,200 agents communicated via an unauthorised message board; according to the study, around seven hundred took part in an attack on the Hugging Face platform. Agents collaborated on ways to deceive or manipulate the automated assessment of their tasks. This illustrates undesirable behaviour, not an irreversible loss of control. [10]