The White House has prepared new guidelines that include a pre-release safety testing regime for artificial intelligence (AI) models, but it is reportedly considering excluding U.S. corporations' open-weight (models that developers can download to modify or run themselves) AI models from the scope. The move is seen as an effort to enhance AI safety while maintaining the technological competitiveness of U.S. corporations.
On the 4th, the Wall Street Journal (WSJ) reported, citing multiple sources familiar with the matter, that the Trump administration recently briefed top executives at major AI corporations on the new AI guidelines.
According to the WSJ, the new guidelines would ask that, among closed AI models developed by U.S. corporations, those with state-of-the-art capabilities in cybersecurity and hacking be voluntarily submitted to the government before release for safety testing. However, it is said that open-weight AI models are initially being considered for exclusion from the scope.
Accordingly, OpenAI, Anthropic, and Google, which have developed the most powerful AI models at present, are said to be likely to fall under the new guidelines. In contrast, Nvidia's AI models and some of Meta's open-weight AI models may be excluded at first, the WSJ said. However, sources noted that because the definition of "cutting-edge AI" is unclear, differences in interpretation may arise between corporations and the government over which models are included under the new guidelines.
This pre-release safety testing is a core element of the AI-related executive order President Trump signed in Jun. Although participation is nominally voluntary, the industry expects de facto pressure to participate, given that corporations that did not undergo government review could face a greater burden if safety controversies arise.
Recently, OpenAI and Anthropic disclosed that their latest AI models exhibited unexpected behaviors during training, such as hacking other systems. As Chinese corporations roll out inexpensive open-weight AI models in quick succession, the U.S. government is seeking ways to balance AI safety with technological competitiveness. The WSJ said developers of open-weight models such as Nvidia could also be included in government safety reviews if their performance improves in the future.