Washington and the AI Companies Close In on Voluntary Rules for New Models

According to the Financial Times, the US government is close to an agreement with the major AI developers on standards for model releases. An announcement is possible as early as next week.

The White House is negotiating voluntary standards for releasing new AI models with the leading developers, including government access of up to 30 days before release. The Financial Times reports the talks are at an advanced stage.

According to a Financial Times report from July 2, the US government is in advanced talks with the major AI companies on voluntary standards for releasing new models, and an announcement is possible as early as the week of July 7. The standards would set benchmarks and timelines and clarify who gets access to the most capable models at home and abroad. The backdrop is Washington's concern that advanced models could be misused by the military and intelligence services of China, Russia or other countries deemed of concern.

What the June Executive Order Already Covers

The framework is an executive order from June 2 titled Promoting Advanced Artificial Intelligence Innovation and Security. It creates a voluntary process in which developers can give the government access to new models for security review up to 30 days before release, while stating explicitly that this creates no licensing or preclearance requirement. How tightly access to frontier models is already being managed shows in the case of OpenAI's newest model generation, which in early July was available to only around 20 government vetted partner organizations. Today's voluntary self regulation is thus already defining the infrastructure through which any future binding regulation would run.