OpenAI's Sol Release Highlights the Murky, Ad Hoc Process for Approving Frontier AI Models
Key Takeaways
- ▸OpenAI's Sol model received approval for public release, but the process by which it was deemed safe remains largely undisclosed to the public and even to government employees and safety experts
- ▸No formal standards, testing protocols, or approval criteria have been publicly established; instead, the process appears ad hoc with private government conversations and select external reviewers
- ▸The Trump administration's forthcoming framework for frontier model evaluation is still incomplete, with cabinet agencies tasked to finalize standards by early August
Summary
OpenAI is rolling out Sol, its latest advanced language model, for public access—a model comparable in capabilities to Anthropic's Fable, which briefly faced a government-imposed moratorium. However, the process by which both models received regulatory clearance remains opaque and poorly defined. Government officials, AI safety experts, and even industry insiders say they don't fully understand how frontier models are evaluated or what criteria determine approval.
The approval process appears to have been ad hoc, involving conversations between OpenAI leadership and government officials like Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross, along with private previews to select users and external evaluators. Yet the specifics of how models are tested, who performed evaluations, and what standards they must meet remain unclear. An executive order published last month outlined a framework for future frontier model evaluations, but the details have not been finalized—six cabinet agencies were instructed to determine a formal process by early August.
The lack of clarity raises concerns about both the robustness of safety assessments and potential conflicts of interest. OpenAI CEO Sam Altman has reportedly offered equity stakes to the Trump administration, and the company's president has been a major donor to Trump's political operation—dynamics that make it difficult for observers to assess whether approvals were based on technical safety grounds or political relationships.
- Industry figures with financial stakes in rapid AI model releases are helping shape AI policy, raising concerns about conflicts of interest and transparent gatekeeping
