White House won't safety-test open-weight AI models, Reuters says

Trump administration advisers told leading AI companies on Tuesday that the federal government will not safety-test open-weight AI models, leaving companies and users to shoulder the risk assessment on systems that can be copied, modified and run outside the lab.

The decision draws a sharp line around Washington’s role in AI oversight. Open-weight models can be downloaded, altered and run by third parties, which makes them valuable to startups, researchers and independent developers, but also easier to adapt for phishing, malware, disinformation and other harmful uses. By declining to run its own tests on that category, the administration is effectively asking the market to police a class of tools that can spread well beyond the companies that built them.

The White House meeting included staff from Meta, Anthropic, Google and OpenAI. A day earlier, a White House official said the administration had finalized a voluntary framework under which companies could submit their most capable frontier models for government safety testing before release. That split approach leaves some room for federal review of the largest closed systems while excluding open-weight models from a direct federal backstop.

The policy line fits a broader debate inside the administration over speed and innovation versus guardrails and safety. The July 2025 AI Action Plan framed AI leadership as a national security imperative and emphasized U.S. technological leadership, reflecting a lighter-touch posture toward regulation. Supporters of that approach argue American firms need room to compete, especially as China pushes hard for its own AI advantage and open-weight systems can circulate quickly around the world.

Critics of the lighter model warn that once weights are public, it becomes far harder to prevent harmful adaptations. That concern reaches beyond cybersecurity into biosecurity, where easier access to powerful models could lower the barrier to dangerous experimentation. If Washington will not safety-test open-weight systems itself, the responsibility shifts to model developers, downstream deployers and the public to catch failures before those tools are reused at scale.

The administration’s separate voluntary testing framework shows it is not rejecting oversight altogether. But the refusal to extend that testing to open-weight models marks a deliberate choice: preserve openness and industry speed, and accept that the most accessible models will move forward without a federal safety check.