Government’s ad hoc process for clearing OpenAI’s Sol model remains undisclosed

OpenAI’s latest large language model, Sol, was released to the public without a publicly disclosed government evaluation process, according to reporting by TechCrunch. The company has not shared details about the government’s review, including who conducted the testing or what criteria were applied. OpenAI did point to external evaluations by organizations such as the U.K. AISI, SecureBio, and Irregular as part of its safety documentation, but the nature of the U.S. government’s involvement remains unclear.

Named officials, including Secretary of Commerce Howard Lutnick, Secretary of the Treasury Scott Bessent, and U.S. National Cyber Director Sean Cairncross, were reported to have had conversations with OpenAI about the model, but the experts involved in testing and the methods used were not identified. OpenAI stated in a late June blog post that it does not believe the current government access process should become the long-term default and that it would work with the government to develop an alternative approach.

An executive order published last month outlined a roadmap for evaluating frontier models but did not specify what kinds of models require scrutiny or which agencies would perform the evaluations. The Department of Commerce’s Center for AI Standards and Innovation appears to be leading the effort, with six cabinet agencies tasked with finalizing a process by early August. However, there is still no agreement on the scope of models subject to government review or the standards to be applied.

Experts quoted in the report expressed concerns about the lack of transparency and the potential for conflicts of interest. Mina Narayanan, a senior research analyst at Georgetown’s Center for Security and Emerging Technology, stated that she lacks visibility into the exact processes used to evaluate frontier models. Andy Konwinski, a computer scientist and co-founder of Databricks, Perplexity, and the Laude Institute, described the process as opaque and questioned who holds the power to make decisions about model releases.

The backdrop to these discussions includes reports of OpenAI offering equity to the administration and political donations, which some critics argue may have influenced the government’s lighter-touch approach to regulating Sol. In contrast, Anthropic’s Fable model faced stricter scrutiny, including a temporary ban on access by foreign nationals due to concerns about jailbreak attempts and government tensions.

Konwinski and others argue that true safety decisions require broader expert involvement, including safety researchers, alignment researchers, interpretability researchers, and data scientists. They propose institutional formats, such as third-party auditing organizations or focused research organizations, to bring more disinterested expertise into the evaluation process. Without such structures, the current secrecy risks undermining public trust in both the AI industry and government oversight.