The United States is in discussions with artificial intelligence companies regarding voluntary standards for new models.

      The US government is in discussions with AI firms regarding voluntary standards related to the release of new models, according to the Financial Times, with an announcement anticipated within the next week. The proposed standards are expected to establish benchmarks and timelines for advanced models and clarify access for users in the US and overseas.

      These negotiations stem from an executive order signed by Trump in June, which requested that developers provide the government with early access to cutting-edge models prior to their broader release. This executive order fell short of enforcing mandatory requirements, and the version that was finalized represented a scaling back from an earlier draft that proposed a much longer review period, as previously reported.

      Achieving this involved months of internal disputes. A three-way power struggle among the Commerce Department, national security officials, and pro-industry aides had previously delayed a more stringent version of the order, after Trump canceled its signing due to concerns it might impede US AI progress.

      According to sources cited by the FT, Washington's primary concern focuses less on consumer safety and more on the potential utility of advanced systems for military or intelligence uses by adversaries such as China and Russia.

      The proposed framework would allow officials to designate a model as a "covered frontier model" and negotiate terms of access prior to its release, without the licensing requirements that the White House has explicitly ruled out. Google is reportedly one of the companies already engaged in discussions, likely regarding an advanced coding model it is developing.

      Neither Google nor the White House has publicly confirmed the particulars of these discussions, and the information is based on FT's sources rather than official statements, meaning the specifics and timing should be treated as tentative until an announcement is made.

      The processes closely follow what was outlined in the June order. Within 60 days of that signing, agencies such as the Treasury Department, the National Security Agency, the Cybersecurity and Infrastructure Security Agency, and the National Institute of Standards and Technology were assigned to create a classified benchmarking process to evaluate the cyber capabilities of advanced models.

      This timeline suggests that early August could be a key period for the framework to become more defined, coinciding with the timeframe mentioned by the FT for an announcement. Several prominent labs, including Microsoft, Google, and xAI, had previously agreed to allow government testing of their models prior to launch, leaving Meta as the last major company resistant to pre-release security reviews.

      OpenAI, too, has encountered a similar situation, having been requested by the administration to delay the unveiling of a subsequent model while reviews were completed. The current discussions appear to aim at transforming the existing disparate processes into a consistent framework with clearly defined benchmarks rather than relying on case-by-case negotiations.

      Earlier this year, Sam Altman conveyed to Congress a preference for funding testing infrastructure over a formal approval system, an inclination that a voluntary standards framework would largely fulfill. This new round of discussions seems to be a way to add operational substance to the June retreat, preventing critiques that the administration simply abandoned oversight.

      A voluntary standard that is accepted by most of the major labs would allow the White House to assert that there is a functioning regime without having to defend a mandatory one in court. Whether these standards will effectively reduce friction or merely formalize the existing informal reviews depends on details not yet revealed in the FT's sourcing or the June order.

      Companies have the option to refuse participation without facing penalties, meaning the framework's effectiveness is contingent upon the willingness of the largest labs to join. At the time of writing, neither the White House nor the companies mentioned in the FT's report had responded promptly to inquiries relayed through wire services. The timeline indicated by the FT, with an announcement expected within the week, suggests that a clearer understanding of the framework's shape should be forthcoming.