Analysis
The Trump administration's new framework for vetting advanced AI models before release is drawing criticism for being both vague and deliberately non-public, according to reporting from Axios and The Verge this week. The voluntary framework defines a 'covered frontier model' as one that is closed-source, state-of-the-art, and poses a national security risk -- but leaves both 'state-of-the-art' and 'national security risk' without clear definitions, giving the administration wide discretion in how and when it applies.
The most consequential design choice is the open-model exemption: the framework explicitly excludes open-weight models, stating nothing in it should be interpreted as restricting them once released. Bloomberg reported separately that China's open-weight models specifically will be spared from the US testing regime as well, adding a geopolitical dimension -- a framework meant to police frontier AI risk that carves out both domestic open-source releases and foreign open-weight models from China.
“That's a meaningful departure from the AISI-style transparent disclosure model the UK has used, most visibly in this week's joint OpenAI/Anthropic incident disclosure.”
Compounding the vagueness, the White House does not plan to make the framework itself public, and it remains unclear which 'trusted partners' would receive early access to advanced models under it, including whether foreign governments could qualify. That's a meaningful departure from the AISI-style transparent disclosure model the UK has used, most visibly in this week's joint OpenAI/Anthropic incident disclosure.
For AI labs, an exemption-heavy, discretionary framework is arguably easier to operate under than binding pre-deployment testing requirements -- but it also does little to reassure enterprise customers or international regulators that US frontier-model oversight has real teeth, at a moment when the UK's AISI has just demonstrated what substantive, disclosed testing looks like in practice.
What to watch: whether Congress or state legislatures move to codify binding AI-testing requirements in response to the framework's vagueness, and whether the administration's non-public approach becomes a point of friction with allies pushing for more transparent AI-safety coordination.
