OMB's "Unbiased AI" Procurement Rule: What Federal AI Vendors and Buyers Need to Know

What the rule is

If you build or buy AI for a federal agency, this is one of the rules that now governs the paperwork. On December 11, 2025, the Office of Management and Budget issued a seven-page memorandum, Memo M-26-04, signed by OMB Director Russell Vought, providing implementation guidance for a July 2025 executive order on the government's use of large language models (LLMs).

This guide is a plain-language walk-through of the mechanics — what agencies must do, what vendors must hand over, and the deadlines — presented as procurement facts, attributed to OMB. It is not a political commentary. If you sit on either side of a federal AI buy, the value is knowing exactly which documents and contract clauses the rule turns into work.

The two principles agencies must apply

OMB frames the executive order as establishing two "unbiased AI principles" that procured LLMs must meet, per Lawfare's read of the memo:

• Truth-seeking — models should emphasize "historical accuracy, scientific inquiry, and objectivity."
• Ideological neutrality — models should function as "neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas."

The scope is LLMs the government procures for its own use. National security systems are exempt from the requirements, though the memo encourages them to comply, and the guidance sunsets after two years. Whatever one thinks of the policy, the operative fact for a contractor is narrow: these two principles are now criteria an agency has to be able to point to when it selects a model.

What vendors actually have to provide

The part that creates real work is the documentation. During procurement, agencies must obtain "sufficient information" to assess whether a model meets the principles. FedScoop reported the required materials include vendor acceptable use policies, training-process summaries, model evaluation scores, product resources or developer guides, and end-user feedback mechanisms.

OMB also sets minimum transparency items agencies must request — acceptable use policies, model or system cards, and a mechanism for users to report outputs that violate the two principles — with enhanced documentation for certain use cases covering system prompts, safety filters, bias evaluations, and any regulatory-compliance adjustments, again per the memo. None of this is exotic for a mature vendor; it is the model-card-and-eval discipline that a serious federal software and AI delivery practice already maintains. But it does have to exist in writing, and it has to be handed to the contracting officer, not just referenced.

The contract language and the deadlines

Two operational details matter most. First, timing: agencies were directed to apply the guidance to all new LLM procurement orders, modify existing contracts "to the extent practicable," and update their procurement policies by March 11, 2026 — a deadline that has already passed, which means these expectations are live in solicitations now, not coming later.

Second, the teeth: FedScoop noted that agencies must "explicitly identify" the compliance factors as conditions for both contract eligibility and termination. In plain terms, meeting the two principles is not a nice-to-have in the evaluation — it can be written in as a gate to win the award and as grounds to end the contract. That elevates the documentation from marketing collateral to contract-defining material.

What to do about it

For vendors, the move is to assemble the packet before the solicitation forces it: a current model or system card, a written acceptable use policy, evaluation results you can stand behind, a training-process summary, and a working channel for users to flag outputs. Have it ready as a document set, because "we can produce that" is slower than "here it is" when a proposal is due.

For agency staff and acquisition officers, the rule turns into checklist items in the evaluation and specific clauses in the contract — which is why the people who read these solicitations closely are exactly the ones who benefit from a plain summary like this. If your role touches federal AI buying, our companion guides on AI for acquisition officers and what changed in federal AI in mid-2026 put this memo in the context of the wider policy landscape. Always confirm the current text of any memo against the primary source before you act on it — guidance evolves, and this one carries a two-year sunset.

Sources: Lawfare — OMB Releases Guidance on the AI Executive Order; FedScoop — OMB requirements for agencies procuring AI; Mintz — AI: The Washington Report, July 2026. Analysis and framing by Precision AI Academy.

Common questions

What is OMB Memo M-26-04? It is a seven-page OMB memorandum issued December 11, 2025 and signed by Director Russell Vought that tells federal agencies how to procure large language models under a July 2025 executive order, requiring models to meet two "unbiased AI" principles.

What are the two "unbiased AI" principles? Truth-seeking — emphasizing historical accuracy, scientific inquiry, and objectivity — and ideological neutrality — functioning as neutral, nonpartisan tools that do not manipulate responses toward ideological dogmas, per OMB.

What documentation do AI vendors have to provide? Per FedScoop, agencies must collect acceptable use policies, training-process summaries, model evaluation scores, product or developer guides, and end-user feedback mechanisms, plus enhanced documentation such as system prompts and bias evaluations for certain use cases.

Is the rule already in effect? Yes. Agencies were directed to update procurement policies by March 11, 2026 and to apply the guidance to new LLM orders, so the expectations appear in solicitations now. The memo sunsets after two years, and national security systems are exempt but encouraged to comply.