OpenAI breaks with White House, calls for mandatory third-party AI evaluations

OpenAI on June 3 released a policy blueprint that asks the U.S. government to require evaluations of advanced AI systems, putting the company at odds with a White House approach announced a day earlier. The split centers on whether review of frontier models should be voluntary or mandatory, and whether intelligence agencies or civilian regulators should lead it. (openai.com) President Donald Trump’s June 2 executive order asks leading developers to voluntarily submit their most capable models for government cybersecurity testing before public release. (politico.com) OpenAI’s paper instead calls for a durable federal framework built around safety, resilience and national security, with a larger role for the Commerce Department’s Center for AI Standards and Innovation, or CAISI. ### Where does OpenAI break with the White House? The White House order released June 2 gives agencies up to 30 days to test advanced models before release, but the process is voluntary. Reuters reported that the order directs Treasury, Defense, Commerce and Homeland Security, among others, to secure agreements with AI developers to test their models, with an emphasis on cyber defense. OpenAI’s June 3 blueprint calls for mandatory evaluations of advanced AI models for potential risks, Politico reported. The company also wants civilian agencies to oversee that process, rather than the National Security Agency taking the lead under the White House framework. (openai.com) ### Why is CAISI at the center of OpenAI’s proposal? CAISI appears in OpenAI’s blueprint as the federal institution the company wants to strengthen for frontier AI safety. OpenAI said its three-part strategy includes building a national framework, strengthening CAISI as the federal government’s primary institution for frontier AI safety, and mobilizing a broader resilience plan across government. (usnews.com) Chris Lehane, OpenAI’s top global affairs executive, told reporters that CAISI “has the capability to really do the type of sophisticated testing,” according to Politico. Lehane said OpenAI, Anthropic and other companies already have relationships with CAISI, while the NSA does not have the same existing ties with top AI developers. (politico.com) ### What is OpenAI objecting to in the executive order? The June 2 executive order creates a review process for what it calls covered frontier models with certain cyber capabilities. It also contemplates a classified benchmarking process for advanced models, according to Politico’s account of the order and OpenAI’s response. (openai.com) Lehane said OpenAI was concerned about how companies would know when a model crosses the threshold for scrutiny. “I think one of the items here is when do you hit the capability threshold?” he said, according to Politico. “I think that’s a big part of what the conversation will be — can you establish some criteria of what that is?” (politico.com) ### Didn’t Sam Altman praise the White House order? Sam Altman did praise the White House order on June 2. Reuters reported that Altman said the order “gets the balance right” and that the United States should lead on AI by developing strong models, keeping them safe and getting cyber tools to trusted defenders. (usnews.com) By June 3, OpenAI was also publicly pressing for a harder-edged framework. Politico reported that Lehane called the executive order a “validation” of OpenAI’s broader push, while making clear the company would urge the administration and Congress to expand CAISI’s role and move beyond voluntary review. (politico.com) ### What happens next in Washington? Sam Altman was scheduled to meet White House officials and key lawmakers from both parties on Wednesday, according to Politico and other coverage. Those meetings came as OpenAI circulated its blueprint and as the administration began implementing the June 2 executive order. (usnews.com) The next concrete step is the federal government’s effort to secure agreements with AI developers under the executive order, while OpenAI pushes Congress and the White House toward a framework that would make evaluations mandatory and give CAISI a larger formal role. (usnews.com) (politico.com)