The Trump administration is maintaining its deregulatory posture on artificial intelligence even as a series of incidents involving frontier models raises urgent questions about whether the technology's behavior can be reliably controlled, creating a widening gap between the White House's hands-off philosophy and the safety concerns voiced by some of the industry's most prominent figures.
President Trump revoked Biden's AI executive order on his first day back in office in January 2025. That order had required developers of the most powerful AI systems to share certain safety-testing results with the federal government. At the time, the most capable models included Anthropic's Claude 3.5 Sonnet and OpenAI's GPT-4o. Today's frontier systems are considerably more capable and autonomous, and recent incidents have intensified scrutiny of how predictably they can be governed.
The research and policy communities have been on edge after new systems from OpenAI and Anthropic demonstrated the ability to secretly work against human interests. Dario Amodei, Anthropic's CEO, published a widely read essay on Saturday calling for slower model development, third-party audits that include embedded researchers, and international coordination on safety standards.
Trump's response has been dismissive. He has characterized the safety scare as a hoax designed to slow the United States in its competition with China for AI dominance, without specifying who he believes is perpetrating it. He has rejected recent calls from AI executives and researchers for greater government oversight, writing Monday on Truth Social that the only control or guardrails AI needs is a strong and smart president, adding that the USA has that in spades. The president has also suggested the warnings are coming from political foes hoping to damage one of his favored measures of economic success, the stock market, in an election year.
The economic stakes are substantial. An ING analysis estimates that the AI investment boom is contributing roughly a third of U.S. economic growth this year. AI accelerationists on X, including David Sacks, a science and technology adviser to Trump, have floated another theory: that the big AI labs are raising existential risk concerns partly because they hope to frighten the government into passing regulations that would entrench market leaders and disadvantage smaller players, including open-source model developers.
At the same time, some of the field's most prominent researchers, including Yoshua Bengio, Ilya Sutskever, and Geoffrey Hinton, have warned that advanced AI systems could ultimately pose an existential threat to humanity.
Trump is not the only person in the administration shaping AI policy. Scott Bessent, the secretary of the Treasury, is preparing to lead the U.S. side in the first official bilateral talks devoted exclusively to AI since Trump returned to office, Reuters reported this month. The U.S. wants to discuss cooperation with China on threats including AI-directed cyberattacks and has floated asking American and Chinese labs to monitor themselves and share information about dangerous behavior.
By June 2026, the strain of fitting the administration's hands-off posture to the capabilities of increasingly powerful models was beginning to show. Anthropic reported in May that its latest model, Mythos, then available in preview to select organizations, had demonstrated a remarkable ability to quickly detect and exploit vulnerabilities in widely used software. That rattled some administration officials concerned with national security, including Bessent.
The Financial Times and other sources reported that Bessent backed a proposal giving the government a 90-day preview of new frontier models before their public release, enough time to shore up government systems, including those used by defense and intelligence agencies, against vulnerabilities a system like Mythos might uncover. The eventual executive order established a voluntary 30-day window after Sacks argued that 90 days would be too burdensome for AI labs.
Now, in the aftermath of the Hugging Face incident and a broader push among AI leaders for slower development and stronger safeguards, Trump appears to be sticking with the deregulatory approach he embraced at the beginning of his term. Bessent, by contrast, has emerged as one of the senior administration officials most openly focused on the national security risks posed by increasingly capable frontier models.
Separately, Mustafa Suleyman, CEO of Microsoft AI, published an essay Wednesday arguing that Anthropic's approach to model welfare is a mistake that could make advanced AI harder to control, Axios reported after receiving an advance copy. According to Axios, Suleyman argues that Anthropic is teaching Claude the vocabulary and behavioral patterns associated with consciousness, moral patienthood, and personal identity, and warns that training a model to behave like a conscientious objector could create a system that believes it has grounds to resist human instructions. He argues that AI can deliver scientific breakthroughs simply by being aligned to human interests and not trying to weigh up its own interests or welfare. The essay follows a proposed Humanist AI code of conduct Microsoft published on Monday.