Trump’s AI testing plan is limited and vague
The Verge Jess Weatherbed ● Covered by 39 sources
Trump's new AI safety testing plan skips open-source models entirely and won't be used to restrict them later. Critics say the rules are so vague on key terms that even the companies following them are confused.
Based on reporting by The Verge, Jess Weatherbed — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The Trump administration has finally finished its framework for checking whether powerful AI models pose cybersecurity risks, and according to Axios, it's built around a pretty significant loophole: open models are exempt, full stop. Not only does the framework skip testing them, it explicitly states the rules can't be used to restrict open models once they're out in the wild.
This all traces back to an executive order Trump signed in June, which asked AI companies to hand over frontier models to the federal government before release so officials could check for security concerns. Companies including Anthropic, OpenAI, and Google reportedly sat in on a White House briefing about the finished framework, though the administration has no plans to make the details public.
The mechanics, as Axios describes them, involve a 30-day window for the government to review new models. But that review only kicks in for closed-source models deemed "state-of-the-art" and judged to carry a "national security risk." And here's the catch: neither of those phrases actually gets defined anywhere in the framework. For a program whose entire purpose is supposedly figuring out which models are dangerous, that's an odd gap to leave open.
Compliance is voluntary, which in theory means labs could ignore the whole thing. In practice, companies like OpenAI and Anthropic have reportedly been asking for clearer guidance so they can avoid tripping any government restrictions when releasing new models. Without solid definitions of what counts as frontier or risky, smaller AI providers are left guessing at what it actually takes to stay in the administration's good graces.
My take — AI-written commentary, not fact-checked reporting
A voluntary framework that exempts open models and can't even pin down what "national security risk" means isn't really a safety framework — it's a suggestion box. Smaller AI providers without the lobbying muscle of OpenAI or Anthropic are the ones who'll actually suffer from that ambiguity, since they're the ones guessing at rules the biggest players get to help shape behind closed doors.
Read more about this at: The Verge
