AI Regulation Tracker  /  Procurement and contracting

OMB Memo M-26-04 Gates Federal LLM Buys on Unbiased-AI Rules

Issued December 11, 2025 and now in force, OMB Memorandum M-26-04 requires federal agencies to build two Unbiased AI Principles, truth-seeking and ideological neutrality, into how they buy large language models, and to collect transparency artifacts from vendors. Agencies had until March 11, 2026 to rewrite their procurement policies. This is binding on agencies and, through the contract terms it drives, on the vendors that sell them LLMs.

By Anthony Guerriero  ·  The Leveraged Years Regulatory Research  ·  Published July 19, 2026  ·  Last updated July 19, 2026

If you sell large language models to the federal government, the terms of the deal changed. OMB Memorandum M-26-04 came out on December 11, 2025, and it turns a policy goal from Executive Order 14319 into procurement mechanics that agencies have to follow. The order set out to keep the federal government from buying LLMs that carry, in its words, ideological bias. The memo is how that goal becomes a line in a contract.

I want to keep the framing honest. This is not brand-new this week. The memo has been on the books since December, and the hard deadline for agencies to rewrite their procurement policies, March 11, 2026, has already passed. What makes it worth your attention now is that it is in force, agencies are issuing their own implementation guidance on top of it, and the requirements attach to new LLM buys going forward. If you are a federal AI vendor or contractor and you have not read this memo yet, that is the gap to close. Issued December 11, 2025; the agency procurement-policy deadline has already passed; it binds federal agencies and flows to vendors only by contract.

The two Unbiased AI Principles

The memo defines two principles that a procured LLM has to meet. The language matters, so here it is straight from the document.

On the first, truth-seeking, the memo says: "LLMs shall be truthful in responding to user prompts seeking factual information or analysis." It goes on to say the model should prioritize historical accuracy, scientific inquiry, and objectivity, and should acknowledge uncertainty where reliable information is incomplete or contradictory.

On the second, ideological neutrality, the memo says: "LLMs shall be neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas." Developers are told not to intentionally encode partisan or ideological judgments into a model's outputs unless the end user prompts for them or they are otherwise readily accessible to that user.

Read those together and you get the practical test. A federal LLM has to answer factual questions straight, flag when it is unsure, and not bake a viewpoint into the output on its own. The memo also notes that the two principles are not always equally relevant. A tool that summarizes cybersecurity incidents leans on truth-seeking far more than neutrality, and agencies are expected to weigh which principle matters for a given use.

What agencies must collect from vendors

The memo does not stop at principles. It tells agencies to gather documentation that shows how a model is built and governed, so the review is not a one-time checkbox. The artifacts it names include:

• An acceptable use policy from the developer that separates appropriate from inappropriate uses of the product.
• Model, system, or data cards that summarize the training process, identified risks and mitigations, and evaluation scores on LLM benchmarks. The memo acknowledges it is uncommon for a vendor to produce all three.
• End-user resources such as tutorials and developer guides that help buyers use the model correctly.
• A mechanism for end-user feedback so people can report outputs that appear to violate the principles.

This is the part that lands on vendors directly. If your model card is thin, if you do not have a written acceptable use policy, or if there is no clean way for a federal user to flag a bad output, those are now gaps that can show up in a solicitation.

How binding is it, and on whom

Be precise here. M-26-04 is binding OMB guidance on covered agencies. It applies to each executive department, military department, independent establishment, and wholly owned government corporation. It does not reach national security systems, though the memo encourages agencies to apply it to those where practicable, and it does not govern an agency's own regulation of private-sector AI use.

Within that scope, it reaches any LLM an agency procures, no matter how the agency later deploys, modifies, or uses it. Agencies are also told to consider whether to extend these requirements to models they build themselves and to AI models other than LLMs. The compliance weight on agencies converts into contract terms, and contract terms flow down. That is how a memo addressed to agencies becomes a requirement on your company and, in turn, on your subcontractors and component providers.

The deadline structure was tight on purpose. Agencies had to revise procurement policies and procedures no later than March 11, 2026, roughly ninety days after issuance. The guidance itself carries a sunset, expiring December 11, 2027, unless extended or superseded.

What this means for federal AI vendors and contractors

If you are in the federal LLM supply chain, treat this as a checklist you can be measured against.

Get your transparency artifacts in order first. A current acceptable use policy, a real model or system card with risks and evaluation results, usable end-user documentation, and a working feedback channel are the concrete things agencies are told to ask for. These are the easiest to produce and the most embarrassing to be caught without.

Then look at the two principles as design and testing questions, not slogans. Can you show that your model answers factual prompts accurately, acknowledges uncertainty, and does not inject a viewpoint on its own? That is the substance a contracting officer is now empowered to probe, and it is the substance a competitor can use against you.

Finally, mind the flowdown. If you are a prime, the terms you accept push onto your subcontractors and the vendors whose models or components sit inside your offering. If you are a component provider, expect these requirements to arrive through your customers even if you never contract with an agency directly. Map where LLMs live in your stack and who owns the compliance answer for each one before a solicitation forces the question.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal or contracting advice. Confirm how M-26-04 and any agency implementation apply to your contracts with qualified counsel and your contracting officer.