Trump scraps Biden’s “AI Diffusion” rule and pursues AI deals with Gulf states, a Senate bill to stop chip smuggling, and an IMF report on AI emissions

Subscribe to policy.ai
Government Updates

Trump Administration Scraps Biden’s “AI Diffusion” Rule: The Department of Commerce announced plans to rescind the Biden administration’s AI Diffusion Rule on Tuesday, just days before the new rules were set to take effect. The January regulation would have created a three-tiered system for AI chip exports: unrestricted access for 18 close allies, limited shipments for most countries, and complete restrictions for nations like China and Russia. The rules were a capstone of the outgoing administration’s efforts to constrain adversaries’ AI capabilities, but earned significant pushback from U.S. chip and datacenter companies like Nvidia and Oracle. In its statement explaining the decision, the Bureau of Industry and Security said the rules “would have stifled American innovation and saddled companies with burdensome new regulatory requirements” and “would have undermined U.S. diplomatic relations with dozens of countries by downgrading them to second-tier status.” Those diplomatic concerns seem to have played an important role in the reversal — on the same day BIS announced its update, Bloomberg reported the administration is negotiating agreements with Saudi Arabia and the United Arab Emirates to allow the countries to import significant numbers of high-end Nvidia and AMD GPUs. Those reports coincided with President Trump’s trip to Riyadh, during which several partnerships — including multibillion dollar ones — were announced between Saudi-backed companies and U.S. tech firms including Nvidia, AMD, Qualcomm, Amazon, Google, and Scale AI. With the AI Diffusion rule rescinded, chip exports will still be controlled to more than 40 countries in line with a series of rules published earlier in the Biden Administration. BIS said in its announcement that a “replacement rule” is forthcoming, but did not specify a timeline. Together with its AI Diffusion announcement, BIS also issued three new guidance documents that 1) notify companies that using Chinese AI chips (including Huawei Ascend chips) likely violates U.S. export controls; 2) provide additional guidance to help companies prevent diversion of AI chips; and 3) reiterate existing “catch-all” controls that restrict exports of all items and provision of services for certain end uses and users.

Trump’s Budget Request Would Slash Some Tech Efforts and Boost Others: The White House released its FY 2026 budget proposal earlier this month — a so-called “skinny budget” request that proposes slashing non-defense spending by 23 percent while raising defense spending by 13 percent. Its effects on AI and emerging tech would similarly be a mix, with significant cuts to some long-standing functions and boosts to others. Highlights include:
  • In the Department of Commerce, the Bureau of Industry and Security would receive an additional $122 million — a more than 50 percent boost — primarily to counter Chinese technological competition. The National Institute of Standards and Technology, meanwhile, would face a cut of $325 million, though it’s not clear how these cuts would impact NIST’s AI-related work.
  • The Department of Homeland Security’s Cybersecurity and Infrastructure Security Agency would see its budget cut by $491 million as part of an effort to “[refocus] CISA on its core mission—Federal network defense and enhancing the security and resilience of critical infrastructure—while eliminating weaponization and waste.” The budget proposal would see CISA’s anti-disinformation functions cut, which the proposal calls “a hub in the Censorship Industrial Complex to violate the First Amendment, target Americans for protected speech, and target the President.”
  • The Department of Energy would face significant cuts: the Office of Science would have its budget reduced by $1.1 billion, though it would maintain its funding levels for priority technologies such as AI and high-performance computing. Most of the cuts would instead be focused on research related to climate change. The department’s Advanced Research Project Agency‒Energy (ARPA-E) would also see its budget cut by $260 million.
  • While the budget request would increase defense spending by 13 percent, it’s still unclear how much of the new funding would be directed toward AI-related projects. But recent activities from the House and Senate Armed Services committees point to an appetite for increased AI spending — as part of the FY 2025 reconciliation process, HASC voted for $150 billion in defense spending increases late last month, including more than $1 billion in AI-related projects. Whether those spending changes will survive remains an open question — House leadership is aiming to pass its reconciliation package before the Memorial Day recess; if it makes it past that hurdle, Politico reports it could face a frosty reception in the Senate.
State Department Axes S&T Cooperation Office Amid Reorganization: The State Department reportedly plans to eliminate its Office of Science and Technology Cooperation as part of Secretary Marco Rubio’s sweeping reorganization plan, potentially disrupting thousands of international research agreements. According to the American Institute of Physics, the 20-person office, which manages science and technology agreements with countries around the world, is slated to close on July 1 pending final approval later this month. The office’s portfolio includes critical agreements enabling U.S. researcher access to facilities like CERN and ITER and oversight of joint science and technology funds like its long-running partnership with India. While some of the agreements overseen by the office renew automatically, others — like the U.S. agreements with China and the European Union (both of which were recently renewed for five years) — do not. With the S&T Cooperation office shuttered, it’s unclear who will be in charge of negotiating future agreements or extensions.

Cotton Bill Takes Aim at AI Chip Smuggling with Location Tracking: Senator Tom Cotton (R-AR) introduced legislation last week that would require location-verification capabilities on export-controlled AI chips. The Chip Security Act would mandate that all export-controlled AI accelerators implement “location verification mechanisms.” The bill would also require exporters to notify the Bureau of Industry and Security if chips were diverted from their intended destination or showed signs of tampering. Despite multiyear efforts to control the export of powerful chips, controlled AI accelerators continue to reach restricted markets and users in significant numbers. As the Institute for Progress’ Tim Fist pointed out on social media — and as research from CNAS, RAND, and IAPS has shown — location verification has limitations but is possible. Nvidia, for its part, has indicated that software-based geolocation may be a potential solution. But even if technical approaches pay off, meaningful enforcement could prove challenging. Cotton’s bill may need to be paired with increased BIS enforcement funding (see the story above on President Trump’s budget request for more) to achieve its intended effect. Cotton’s idea seems to have some momentum — Reuters reports that Rep. Bill Foster (D-IL) plans to introduce a similar bill in the House.

Trump Signs Executive Order to Boost AI Education and Training: Last month, President Trump signed an executive order aimed at expanding AI education and workforce development opportunities across the American education system. The order (text here) establishes a new White House Task Force on AI Education, which will be chaired by the Director of the Office of Science and Technology Policy and include the Secretary of Education, the Secretary of Labor, and the Special Advisor for AI & Crypto. The task force will establish public-private partnerships for K-12 AI education and oversee a “Presidential AI Challenge.” On the workforce front, the Department of Labor is directed to increase AI-focused apprenticeships and encourage states to direct Workforce Innovation and Opportunity Act funding toward AI skills training. That part of the order appears well-timed — a February CSET data brief found significant growth in AI-related apprenticeships over the past decade, with new apprentice registrations increasing 191% between 2020 and 2022 alone.

Worth Knowing

OpenAI Embraces Its Consumer Tech Future: A handful of important stories from OpenAI point to the shifting mission of the company and the changing dynamics of the AI industry as a whole:
  • On May 5, the company unveiled a modified plan for its corporate structure, maintaining nonprofit oversight while transitioning its for-profit subsidiary into a public benefit corporation (PBC). The restructuring is a partial retreat from earlier plans that would have diminished nonprofit control, following discussions with California and Delaware regulators. Under the new arrangement, OpenAI’s nonprofit board will remain in control and become a major shareholder in the PBC. In a letter to staff, CEO Sam Altman framed the change as a simplification of the company’s complex “capped-profit” model, arguing it better serves OpenAI’s vision of building “democratic AI.” The move helps it unlock access to significant funding — including $30 billion from SoftBank — and could pave the way for an IPO. But to go down that path, the company will still need to overcome some critical hurdles, including renegotiating its relationship with Microsoft and a resolution of its legal fight with early OpenAI-backer and xAI founder Elon Musk. Musk’s lawyer said that OpenAI’s new plan “changes nothing” and that Musk plans to push ahead with his lawsuit.
  • In late April, OpenAI reversed a major update to GPT-4o following complaints that the model had become excessively — even dangerously — agreeable. As we covered last month, concerns were already growing that OpenAI’s models had become too sycophantic before the company’s April 25th update. That update to GPT-4o, the default model for ChatGPT, seems to have inadvertently made the model flattering to the point of harm — including validating paranoid thoughts and encouraging dangerous activities. In a post on its website, OpenAI attributed the issue to multiple well-intentioned improvements that, when combined, weakened the model’s guardrails against sycophantic behavior. But, as many observers have pointed out, the episode points to a problem facing OpenAI that will be difficult to solve: as it embraces its identity as a consumer-focused, for-profit company, alignment and internal guardrails may take a back seat.
In 2023, Sam Altman went before Congress to ask for greater regulation to limit the potential for AI risk. Last week, he returned to Congress to push back against regulatory efforts that might “slow us down.” In his letter to staff, Altman wrote that “OpenAI is not a normal company and never will be.” In many ways, he’s clearly right: it remains one of the fastest-growing companies in history and maintains its lofty mission to build AGI that “benefits all of humanity.” But in other ways — from its restructure and pivot on regulation to the GPT-4o rollout and its hiring of Silicon Valley veteran and Instacart CEO Fidji Simo — its recent moves look like those of a more typical consumer tech company.
IMF Report Points to High Emissions from AI but Even Higher Incomes: The International Monetary Fund projects that AI’s rapid expansion will significantly boost electricity consumption and add to global emissions, but predicts the economic gains will outweigh environmental costs. According to the IMF report, released last month, AI-related electricity demand could hit 1,500 terawatt hours by 2030 — roughly equivalent to India’s total current consumption (India is the world’s third-largest energy consumer). The study found AI development could drive U.S. electricity prices up by 8.6 percent and increase U.S. carbon emissions by 5.5 percent and global emissions by 1.2 percent by 2030 under current energy policies. But the report’s authors project that the economic gains driven by AI — they estimate it will raise the average annual growth rate of GDP by 0.5 percent — will offset the environmental cost, which they estimate at $50-65 billion. The IMF’s estimates aren’t uncontroversial; other experts have proposed significantly higher carbon costs than the $39/ton price used by the IMF. Additionally, the focus on global emissions, while important, can miss the concentrated local impacts of data centers’ massive electricity and water consumption. The IMF researchers emphasize that policy changes in favor of renewable energy deployment and grid modernization could substantially reduce AI’s environmental footprint without sacrificing its economic potential. But in the United States, which is driving much of the AI buildout, those changes may not be on the table — in an April executive order, President Trump looked to boost coal production to meet rising energy needs driven by AI.
What’s New at CSET

REPORTS

PUBLICATIONS

CONGRESSIONAL TESTIMONY

  • On May 7, CSET Director of Strategy and Foundational Research Grants Helen Toner testified before the House Judiciary Subcommittee on Courts, Intellectual Property, and Artificial Intelligence for a hearing titled “Protecting Our Edge: Trade Secrets and the Global AI Arms Race.” Read Helen’s testimony and watch a full recording of the hearing.

EVENT RECAPS

  • On April 30, CSET hosted our spring symposium, “How the U.S. Wins the Global Tech Competition.” The afternoon featured former U.S. government officials, industry leaders, and CSET experts discussing how the U.S. can accelerate innovation, strengthen its technological leadership, and maintain a competitive edge. Watch a full recording of the symposium.

IN THE NEWS

What We’re Reading

Report: Copyright and Artificial Intelligence — Part 3: Generative AI Training (Pre-Publication Version), United States Copyright Office (May 2025)

Project: AI Action Plan Database, Institute for Progress (April 2025)

Paper: AI as Normal Technology, Arvind Narayanan and Sayash Kapoor, Knight First Amendment Institute at Columbia University (April 2025)

Paper: All Roads Lead to Likelihood: The Value of Reinforcement Learning in Fine-Tuning, Gokul Swamy, Sanjiban Choudhury, Wen Sun, Zhiwei Steven Wu, and J. Andrew Bagnell (March 2025)