Trump makes AI model sharing voluntary

President Donald Trump’s administration is preparing an executive order that would make it voluntary for AI companies to share advanced models and technical details with federal agencies before release, according to reports by Axios and Gizmodo. The draft order would replace a harder-edged approach under discussion earlier this month with an opt-in framework for government review, the reports said. The White House has not publicly released the text. The proposal comes as U.S. officials debate how to oversee frontier AI systems without imposing mandatory disclosures on companies such as OpenAI, Anthropic and Google. ### Which requirement is the draft order changing? The Biden administration’s October 30, 2023 executive order required developers of the most powerful AI systems to provide safety test results and other information to the federal government when their models met thresholds that could pose serious national security, economic or public health risks. The order directed agencies to use authorities under the Defense Production Act to obtain that information, according to the White House fact sheet and the published text of Executive Order 14110. Axios reported on May 20 that the Trump draft now outlines a voluntary framework for AI developers to inform the government about new releases, and Gizmodo, citing Axios, said the most significant safeguard in the draft may no longer be mandatory. Bloomberg reported on May 8 that an earlier draft had already omitted mandatory model tests. ### What would companies actually be asked to do? (govinfo.gov) Politico reported on May 20 that the directive would ask technology companies to submit advanced AI models for review by federal agencies, citing people familiar with the draft. Nextgov said the expected order could give the National Security Agency a role in voluntary testing of some AI models, particularly cyber-focused systems. (axios.com) The reports indicate companies would be able to opt into sharing models or technical information rather than face a compulsory handover. International Business Times, summarizing the draft, said the framework would facilitate pre-deployment safety testing with government agencies. The White House has not published a formal participant list or compliance process. (politico.com) ### Why does this matter for federal oversight? The change would narrow the federal government’s guaranteed access to frontier model information because agencies would depend on company participation rather than a mandatory reporting trigger. That is a direct contrast with the structure of Executive Order 14110, which paired safety reporting with specific thresholds and federal authority to demand information. (ibtimes.sg) The reporting also suggests the administration is trying to balance pressure from national security officials seeking earlier visibility into advanced models with opposition from technology companies to compulsory disclosures. The Economic Times, citing the emerging draft, said the order aims to establish a voluntary framework before public release of new models. (govinfo.gov) ### How does the U.S. approach compare with Europe’s this week? European Union governments and European Parliament negotiators reached a provisional agreement on May 7 to simplify parts of the bloc’s AI Act, according to the Council of the European Union and Euronews. The changes were presented as a streamlining exercise, not a repeal of the law’s core risk-based structure. (economictimes.indiatimes.com) Euronews reported that the EU package extends some deadlines, narrows some obligations and reshapes enforcement, while preserving the framework for high-risk systems. That leaves Europe still operating through binding legal obligations even as it trims complexity, unlike the U.S. draft reported this week, which would make at least one key channel of model sharing optional. (consilium.europa.eu) ### What happens next? Axios and Nextgov both reported the White House could release the executive order as soon as this week. Until the administration publishes the text, key details remain unresolved, including which agencies would receive model information, what classes of models would qualify and whether any reporting thresholds would survive in another form. (axios.com) (euronews.com)