According to Axios, the U.S. government said Monday local time that it had completed, on schedule, work on a voluntary framework for evaluating advanced AI models. However, the White House did not disclose what the framework specifically includes, which agencies or companies have reviewed it, or when companies will begin adopting it.
Policymakers, AI safety advocates, and U.S. allies are waiting to learn what regulatory rules will govern the world's most powerful AI models. The White House said: “The voluntary framework proposed in the June 2 executive order was completed before the deadline.”
A White House official said: “We are currently discussing next steps with the industry.” The official added that the government was engaging with “many more industry partners than just Anthropic, OpenAI, and Google.”
Before the deadline, the three AI labs had already provided the government with feedback on the draft framework.
The framework is intended to provide AI developers with a structured process to help them work with the government to determine whether models under development fall within the scope of regulation.
According to reports, the framework should clearly specify the confidentiality, cybersecurity, internal risk management, intellectual property protection, and nondisclosure requirements that must be followed when the government obtains a relevant model up to 30 days before its release.
In addition, the framework should clarify which “trusted partners” may receive early access to models.
A U.S. presidential executive order explicitly stipulates that the benchmarking process used to evaluate an AI model's advanced cyberattack capabilities will be classified. The threshold for models covered by the executive order is also classified information and will be disclosed to AI developers and researchers only “as appropriate.”
However, the executive order did not classify the voluntary framework itself, so policymakers and other observers had previously expected to see the relevant details.
AI companies want clear guidance as soon as possible so they can determine whether models under development might be subject to the framework. A White House official said: “Even if certain information is not classified, that does not mean we will disseminate it publicly to everyone.”
According to people familiar with the matter, the White House plans to hold a staff-level meeting with relevant companies on Tuesday local time to review the framework.