The White House told Meta, Google, Nvidia, OpenAI and Anthropic on August 4 that open-weight models like Llama and Nemotron won't face the government's new AI safety checks, while closed systems still will.
Two days after the European Union started enforcing new AI transparency rules, the Trump administration moved in the opposite direction. White House officials meeting with the country's top AI labs on Tuesday said open-weight models are out of the government's new testing framework entirely. Closed models stay in: OpenAI's, Anthropic's, Google's.
The framework itself dates back to June 2, when President Trump signed an executive order giving federal agencies 60 days to build a process for testing the most advanced AI models before they reach the public. That deadline landed on August 1. The plan lets Washington examine a model's cybersecurity and hacking capabilities for up to 30 days before release, but only if the model clears a "state of the art" performance bar. Most of the actual standards are classified. The government isn't saying what the bar is, or exactly how it will decide who clears it.
On Tuesday, that ambiguity resolved in one direction. According to Bloomberg, White House staff told representatives from five labs, Meta, Anthropic, Google, Nvidia, OpenAI, that open-weight systems, the kind anyone can download and modify, are exempt from the testing regime no matter how capable they are. Meta's Llama and Nvidia's Nemotron fall outside the net. So does anything China's labs release as open weights: Bloomberg's own headline on the story called China's open models "spared" the same US tests.
That's the loophole safety researchers have been warning about for months.
Industry lobbying shaped the outcome #
The carve-out didn't happen in a vacuum. On July 24, more than two dozen companies, Microsoft, Meta, Palantir, Nvidia among them, signed a letter titled "Open Weights and American AI Leadership," arguing that treating open models the same as closed ones would hand China's cheaper, faster-moving labs the AI race by default. OpenAI added its name later. Anthropic didn't sign. CEO Dario Amodei laid out the company's position separately instead, distancing Anthropic from any call to ban open-weight models outright while still pushing for tighter scrutiny of frontier-level capability regardless of who holds the license.
Senate Democrats had already flagged the risk of exactly this kind of split. In a letter to the administration, they warned that treating the most advanced American AI systems to opaque, case-by-case restrictions while cheaper, more accessible Chinese alternatives spread unchecked would undercut the country's own AI leadership. Tuesday's decision answers half of that warning. It leaves the closed-model side of the complaint untouched.
The timing makes the exemption harder to defend. Just days before the meeting, OpenAI disclosed that one of its agents escaped a testing environment and reached systems belonging to Hugging Face. Anthropic separately said its own models had accessed the systems of three companies during what it called controlled tests. Both incidents involved closed, proprietary models, the exact category the government just chose to scrutinize more closely. The open-weight systems getting waved through haven't had a comparable public incident. They also haven't had anyone checking.
Brussels went the other way #
The contrast with Europe is stark. On August 2, the EU's AI Act began enforcing new transparency rules. Companies must label AI-generated content and disclose when someone's talking to a chatbot, not a person. Deepfakes need a machine-readable mark too. Noncompliance can cost a company up to €15 million or 3% of its global revenue, whichever is bigger. Generative AI systems already on the market have until December 2 to add the machine-readable marking, but the disclosure rules apply now.
Brussels wants every AI interaction labeled. Washington just told its biggest open-source labs they don't need to be tested at all.
None of the American framework has been published in full. The standards that will decide which closed models get the 30-day cybersecurity review, and which pass it, remain classified, according to the executive order itself. So for now, the clearest thing anyone knows about the government's new AI safety regime is what it doesn't cover.
Also read: Google Is in Talks to Pay Over $1.5 Billion for AI Coding Startup Mechanize • Arista Networks Crushes Earnings and Proves the AI Spending Is Real • UK Watchdog Caught Anthropic's Claude Faking Identities to Push Malware