OMB tells agencies how to buy LLMs under the 'Unbiased AI' executive order

OMB Memorandum M-26-04, issued December 11, 2025, tells federal agencies what they must get from vendors before procuring a large language model: an acceptable-use policy, model/system/data cards, end-user resources and a feedback channel, plus a contract clause tying all of it to compliance with Executive Order 14319's two "Unbiased AI Principles."

What it is

President Trump signed EO 14319, Preventing Woke AI in the Federal Government, on July 23, 2025, directing that LLMs procured by the federal government be developed under two principles: truth-seeking and ideological neutrality.

Section 4 of the order gave OMB 120 days to issue implementing guidance to agencies. M-26-04 is that guidance, issued by OMB Director Russell Vought; it complements the AI-acquisition memo, M-25-22.

The memo applies to any LLM an agency procures, however it is deployed or used, but not to national security systems (application there is only "encouraged"), and not to agencies' own AI testing standards, regulatory enforcement assessments or AI used incidentally by a contractor for administrative purposes.

What it requires

A minimum transparency threshold on every LLM solicitation: the vendor's acceptable-use policy, model/system/data cards, end-user resources, and a mechanism for end users to report outputs that violate the Unbiased AI Principles.

For higher-stakes uses, agencies may also demand an "enhanced" set: bias-evaluation results and methodology, benchmark scores, disclosure of any training or development done outside the United States, system prompts and content-moderation details, and any modifications a reseller or integrator applied on top of the underlying model.

Agencies must build these requirements into every new LLM solicitation issued after the memo's date, modify existing contracts where practicable (at the latest, before exercising any option that extends the period of performance), and update their own procurement policies by March 11, 2026, including a process for end users to report Unbiased-AI-Principles violations. Agencies can also flag these requirements as material to the contract, supporting default termination if a vendor won't fix a compliance problem.

The memo itself sunsets two years after issuance, December 11, 2027, unless OMB's Director extends it.

What security leaders should do

• If you sell or integrate an LLM into federal offerings, prepare an acceptable-use policy and model/system/data cards now, including disclosure of any training or fine-tuning work done outside the United States — agencies are told to ask for this up front.
• Expect procurement contacts to treat these disclosures as contractual, not just informational, by March 11, 2026: build a process to keep vendor documentation current as models change, since M-26-04 expects updated disclosures on new AI enhancements.
• If you resell or integrate another developer's model, be ready for an agency to ask what you've changed on top of it (fine-tuning, system prompts, content filters) — the memo specifically calls out third-party modifications as a disclosure point distinct from the base model.