In recent months the U.S. regulatory scene for artificial intelligence has accelerated, and informal consultations between the government and leading AI labs produced visible results: OpenAI and Anthropic obtained government nods or clearance-like feedback before broadly releasing their most powerful models. A few months ago, that outcome would have been unlikely.
Why it matters
Those behind-the-scenes negotiations exposed how the absence of settled regulatory frameworks and limited technical capacity forced both government and industry into ad hoc arrangements. The process has involved threats of export controls, licensing requirements and negotiations across multiple agencies that sometimes take conflicting positions.
Background and contradictions
A Biden-era AI executive order required companies to share safety testing results with the government, including whether models could be manipulated to bypass built-in safety guardrails (so-called ‘‘jailbreaking’’ or safety bypassing). That vulnerability was highlighted by Amazon last month and helped prompt export control measures affecting Anthropic.
The Trump administration revoked the Biden-era reporting requirements, yet in specific cases authorities have still exerted pressure on industry over safety concerns. When tensions over Anthropic peaked, there was no unified framework to determine how severe a jailbreak or safety bypass needed to be to trigger enforcement — a person familiar with the situation told Axios that a standardized assessment system might have prevented export controls.
Criticism: expertise and funding shortfalls
Former Biden technology official Asad Ramzanali noted that other countries have adopted privacy laws, updated competition rules and passed transparency and research access measures, while the U.S. has not made comparable moves. "We didn't do any of it," he said, arguing the country lacked a strong foundation to craft AI rules now.
Kevin Frazier of the Cato Institute and director of the University of Texas' AI and law program argued the government never recruited or retained sufficient technical expertise: fewer than 1 percent of AI Ph.D. holders enter government roles. Agencies such as the Cybersecurity and Infrastructure Security Agency (CISA) and the Center for AI Standards and Innovation (CAISI) have also been sidelined and underfunded, he said. CAISI's operating budget is currently $15 million but, according to the Institute for Progress, it needs roughly $84 million annually to meet the Trump administration's AI action-plan requirements.
Frazier emphasized there is no single right answer for governing AI, but many wrong ones — including weak technical capacity and failure to build public trust. He suggested the federal government could have learned from state-level initiatives such as Oklahoma's free AI literacy training, Utah's regulatory sandboxes for supervised testing, and Massachusetts' data-practice efforts.
Congress has not passed comprehensive AI safety legislation, despite years of bipartisan attention and public interest, leaving gaps filled by one-off regulatory fixes.
Rep. Josh Gottheimer (D-N.J.) told Axios, "Right now, there is far too much confusion with the White House's AI vetting process — both for the country and for our leading AI developers."
Industry response
OpenAI CEO Sam Altman told CNBC that the negotiation process around the wide release of GPT 5.6 was "very productive." He said it was the company's first time going through that level of government review and that they would learn how to improve the process for next time.
Industry leaders have stressed that companies need to be confident in their safety claims because if those claims don't hold up, public and regulatory discomfort can escalate quickly. The White House said on Tuesday that it neither approved nor disapproved OpenAI's decision to release the model.
What to watch next
The administration and industry are continuing to work on the voluntary framework required by the June AI executive order; the framework is due August 1 per the order. Meanwhile, debates about export controls, the handling of jailbreak vulnerabilities and the limited pool of technical expertise in government suggest that without stronger institutional and legal foundations, future disputes may again be resolved through case-by-case pressure rather than predictable rules.
