AI Accelerationists Take the State: Trump’s OMB AI Memo
With its recently released 2025 OMB Memo on federal use and procurement of AI, Driving Efficient Acquisition of Artificial Intelligence in Government (M-25–22), the Trump Administration all but erases the Biden Administration’s 2024 predecessor OMB Memo, Advancing the Responsible Acquisition of Artificial Intelligence in Government (M-24–18).
Gone is any discussion of “responsibility” and the previously adopted testing requirements for “rights-impacting” and “safety-impacting” AI systems. Not surprisingly, gone are mentions of “equity” and “bias” and “environmental.” The Trump OMB Memo basically picks up where Trump 1 left off with AI policy, as if the intervening four years of AI policymaking and state capacity building had not happened — with one notable exception. The federal agency Chief AI Officers — a post that the Biden OMB Memo created — remain, although their duties have shifted from managing risk to speeding acquisition.
I’ll get to some specifics after this general observation: The difference between the two OMB approaches is not just the U-turn on AI risk and harms. The Biden Administration adopted the theory of “procurement as policy” (Deirdre Mulligan & Ken Bamberger). Its OMB Memo sought to use the federal procurement power to shape the market for more trustworthy AI. If the government leveraged its purchases (as well as grants) to demand risk assessments, transparency, risk mitigation, and ongoing adversarial testing from its vendors, then those practices would become standard throughout the market for AI products and services. The Trump OMB Memo evinces very little interest in using procurement as policy. Given that OMB Director Russell Vought wants to make the federal government small enough to drown in a bathtub (approximating Grover Norquist’s quote), he is not too interested in policy. Rather, the focus of his OMB Memo is almost exclusively on deploying AI systems faster, cheaper, and with favorable terms around IP and data. Still, there is a bit of policymaking in the Trump OMB Memo, maybe incidentally. It really seems to want to prevent vendors from locking the government into particular proprietary systems. To support competition, it pushes agencies to favor open model weight systems, open APIs, and interoperability and portability in system architectures. If most of the new OMB Memo is a rejection of what came before, this one piece is an amplification.
NIST and Consensus Standard-Setting are Gone
Biden OMB Memo: It had required alignment with NIST’s 2024 AI Risk Management Framework to ensure standardized risk assessment. The memo had also lifted up international standards, the formation of which NIST plays (or played) a leadership role.
Trump OMB Memo: There is no requirement to follow NIST’s AI RMF; agencies are free to come up with their own performance standards. The whole emphasis is on performance, rather than risk. Performance standards are surely needed, given how much AI doesn’t actually work (Arvind Narayanan). But the sidelining of NIST is another example of this Administration’s squandering American soft power and state capacity. What NIST has to say about AI sociotechnical governance impacts the whole free world. The Trump OMB Memo ignores it.
Equity, Civil Rights, Environmental Considerations Are Gone
Biden OMB Memo: The list of requirements for “rights-impacting” AI systems was long. In short, agencies had to ensure that vendors conducted risk and impact assessments and evaluations on an ongoing basis, conducted adversarial testing and red-teaming, and complied with various transparency and notice provisions. Many of these values were set forth in the 2022 Blueprint for an AI Bill of Rights and in the 2023 Executive Order on AI. The Memo had also acknowledged that high energy consumption for compute was a consideration that agencies should take into account in making procurement decisions. .
Trump OMB Memo: It mentions civil rights compliance, and expects agencies to reserve the right and have access to conduct evaluations as necessary. But that’s about it. It replaces the “rights-impacting” and “safety-impacting” typology of the Biden OMB Memo with the single term “high-impact AI.” Agencies are told to determine as best they can whether it’s a high-impact AI that they’re seeking and then let vendors know that they will have to meet additional transparency and impact assessment requirements. The term refers to:
“AI with an output that serves as the primary basis for decisions or actions with legal, material, binding, or significant effect on: an individual or entity’s civil rights, civil liberties, or privacy; or an individual or entity’s access to education, housing, insurance, credit, employment, and other programs; or an individual or entity’s access to critical government resources or services; or human life, well-being; or critical infrastructure or public safety; or strategic assets or resources, including high-value property and information marked as sensitive or classified by the Federal Government.”
Generative AI Guidance is Gone
Biden OMB Memo: It had required agencies to make sure the generative AI systems they procured transmitted watermarks, metadata, and other provenance markers on possibly deceptive synthetic content. Those systems also had to provide information about training data, data labor, compute, model architecture, and relevant evaluations. And they would be contractually bound to make best efforts to filter out CSAM, NCII, and other toxic content.
Trump OMB Memo: It will develop “playbooks” on GenAI, but otherwise doesn’t deal with it.
Implications for Federal AI Deployment and the AI Market
If the Trump OMB Memo is implemented, we can expect agencies to move fast in acquiring AI systems. They will be focused on cost and effectiveness, but not much on safety or harms. There may be a positive spillover from the OMB’s ostensible commitment to a competitive AI marketplace. If the federal government really favors smaller players, open model weights, and modular systems that pose less risk of lock-in, maybe new entrants will have a shot at big federal contracts. Among the many likely negative spillovers is the thing that won’t happen: a robust ecosystem of AI measurement and evaluation methods and standards. There’s widespread agreement that this has to get a lot better if buyers are going to be able to compare systems and the public can trust that they work well and fairly. (Mathew Burtell & Helen Toner). The Biden OMB Memo sought to turbocharge that development through federal demand. The Trump OMB Memo doesn’t ask for it and seems to view that kind of science as an obstacle to AI adoption.
The next shoe to drop will be the Trump Administration’s recission of Biden’s National Security AI Memo. There too, the Biden Administration had sought to use the massive heft of federal demand to shape the market for trustworthy AI. Undoubtedly, the revision will trash that approach. There too, I suspect, the accelerationists will have their way.
