OpenAI's new blueprint calls for federal AI safety framework with CAISI evaluations

OpenAI has published a new policy document titled 'Democratic Governance of Frontier AI: A Blueprint For A Federal Framework,' released shortly after a new Executive Order on AI. The blueprint proposes a federal approach to governing frontier AI systems, with a central role for CAISI (Civil AI Safety Institute). Crucially, OpenAI explicitly states that CAISI's role should be limited to conducting evaluations and recommending mitigations—not approving or blocking deployments. This distinguishes it from a licensing regime. The document also calls for 'meaningful accountability mechanisms' for companies that fail to comply with safety obligations, though it remains vague on consequences for dangerous models. It advocates for federal preemption of state laws like California's SB 53, arguing that a unified national framework is necessary.

The blueprint places heavy emphasis on recursive self-improvement (RSI), which OpenAI describes as 'potentially the most consequential frontier safety issue of the coming decade.' It urges CAISI to treat RSI as an urgent priority and to develop standards for independent technical assessments that give policymakers ongoing visibility into progress toward RSI. The document also praises the US's unique position to lead, calls for maintaining compute advantage, and supports empowering CAISI. Sam Altman is currently in Washington, DC meeting with Speaker Mike Johnson and Minority Leader Hakeem Jeffries to discuss the proposal. Commenters have noted that the blueprint's non-classified evaluation process avoids the risks of a de facto licensing system seen in the Trump-era EO.