The rush to get government approval before releasing powerful AI models would have seemed absurd just months ago. Now it is the new normal. OpenAI and Anthropic both cleared their latest systems with Washington before going public, a striking shift in how the world's leading AI labs operate under the Trump administration's approach to oversight.
The scramble reflects a deeper problem that experts say did not have to happen. The U.S. has spent years without a coherent AI governance framework, leaving companies and agencies to improvise as each new powerful model gets built. The result is a patchwork of export controls, licensing demands, and negotiations across a fractured government apparatus that sometimes works at cross purposes.
One flashpoint illustrates the dysfunction. When Amazon security researchers flagged vulnerabilities in Anthropic's latest model, the company faced export controls. But there was no standardized definition of what severity of safety flaw should trigger government action. Had Washington and industry agreed on thresholds beforehand, the export controls might have been avoided entirely, according to people close to the situation.
The Trump administration scrapped the Biden-era requirement that companies hand over safety testing data to the government. Yet companies still feel heavy pressure to cooperate, knowing what happened to Anthropic. OpenAI CEO Sam Altman described his latest government submission as smooth on cable news, but also acknowledged his team is already planning how to improve the process next time. That language suggests less confidence than it projects.
Congress has failed for years to pass any comprehensive AI safety law that might have prevented this ad hoc firefighting. Meanwhile, the government has not invested in the technical expertise it needs to regulate the industry properly. Fewer than 1 percent of AI Ph.D.s work for the federal government, and agencies tasked with setting standards remain underfunded and sidelined.
The Center for AI Standards and Innovation operates on $15 million annually but needs $84 million to execute Trump's AI action plan, according to one estimate. CISA and related bodies have lacked both resources and authority to build coherent rules from the start. The government could have learned from state experiments like Utah's regulatory sandboxes or Oklahoma's AI literacy programs, but Washington never built on those models.
A former Biden administration technology official noted that the rest of the world moved forward with tech privacy laws, updated antitrust rules, and transparency requirements. The U.S. did none of it, squandering years that could have created a foundation for smarter AI governance today. What is happening now is necessary, he said, but it should never have been necessary at all.
Rep. Josh Gottheimer voiced frustration about the current vetting confusion, saying there is far too much opacity in how the White House evaluates powerful models. The Trump administration says its cyber security executive order provisions are voluntary and that it has not invoked the Defense Production Act the way Biden did. But the practical reality is that AI companies know the government is watching closely, and they know what happened to competitors who got it wrong.
The voluntary framework both sides agreed to work on is due August 1. Whether that produces something useful or just another layer of bureaucratic theater remains to be seen. What is clear is that the AI industry is no longer moving at its own pace. It is moving at the government's pace, however haphazardly that pace is being set.
Author James Rodriguez: "This was avoidable if Washington had done its homework years ago instead of letting AI companies write their own rules."
Comments