OMB M-26-04 and the Unbiased AI Principles for federal LLM purchases, explained

OMB Memorandum M-26-04 implements Executive Order 14319: agencies buying large language models must write truth-seeking and ideological neutrality into contracts, request a minimum set of vendor documentation, and had to update procurement policy by March 11, 2026.

OMB Memorandum M-26-04 turns two principles from an executive order into contract terms for every large language model a federal agency buys. Most of its weight falls on documentation: what agencies must ask vendors for, what they may ask for, and how the answers become enforceable. This explainer covers four parts: where the memo comes from, what it covers, what agencies must do and by when, and what vendors are asked to provide.

Where the memo comes from

Executive Order 14319, Preventing Woke AI in the Federal Government, was signed on July 23, 2025 and published in the Federal Register on July 28, 2025 [1]. It directs agency heads, consistent with applicable law and OMB guidance, to procure only large language models (LLMs) developed in line with two “Unbiased AI Principles” [1]:

  • Truth-seeking. LLMs are truthful in responding to prompts seeking factual information or analysis, prioritize historical accuracy, scientific inquiry and objectivity, and acknowledge uncertainty where reliable information is incomplete or contradictory.
  • Ideological neutrality. LLMs are neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas. The order names diversity, equity and inclusion as one. Developers do not intentionally encode partisan or ideological judgments into outputs unless the user prompts for them or they are readily accessible to the user.

The order defines an LLM as a generative AI model trained on vast, diverse datasets that enable it to generate natural-language responses to user prompts [1]. It told the OMB Director to issue implementing guidance within 120 days. That guidance had to account for technical limitations, let vendors show compliance with the neutrality principle by disclosing system prompts, specifications, evaluations or other documentation, avoid requiring disclosure of model weights where practicable, avoid over-prescription, and make exceptions for national security systems [1].

OMB issued M-26-04, Increasing Public Trust in Artificial Intelligence Through Unbiased AI Principles, on December 11, 2025 [2]. It says it complements M-25-22 on AI acquisition [2]. As checked on November 14, 2026, OMB’s memoranda page lists it among the 2026 memoranda [3].

One discrepancy in the memo itself: its first footnote cites the order as “Executive Order 14139”, while its text and the Federal Register copy of the order use 14319 [2] [1]. We use 14319, the number in the Federal Register.

What it covers

The memo applies to executive departments, military departments, independent establishments and wholly owned government corporations. It applies to any LLM an agency procures, however the agency deploys, modifies or uses it [2]. National security systems are excluded, though applying the memo to them is encouraged where practicable [2]. Agencies need not apply it to LLMs acquired under a free, open-source license, but should set up due diligence procedures to check such models against the principles and against M-25-21 before use [2].

Agencies must consider the factors in the memo’s appendix in deciding whether to apply the requirements to LLMs they develop themselves and to AI models other than LLMs that they procure [2]. The memo does not govern [2]:

  • regulatory actions about non-agency uses of AI;
  • assessments of an AI application because its provider is the target of an enforcement, law enforcement or national security action, or because a criminal suspect used it;
  • development of metrics, methods and standards for testing AI for general use rather than for one agency application; and
  • AI a contractor uses incidentally for administrative purposes, when the agency did not direct it and the contract does not need it.

What agencies must do, and by when

The memo sets four duties [2]:

  • New solicitations. Any solicitation or order for an LLM issued after the memo’s date must include contract requirements on compliance with the principles, and so must solicitations for other AI models the agency decides to cover.
  • Existing contracts. Agencies should, to the extent practicable, modify existing LLM contracts, at the latest before exercising an option that extends the period of performance.
  • Procurement policy. No later than March 11, 2026, agencies must update their policies and procedures so that LLM contracts include these requirements, with a process for agency users to report outputs that violate the principles.
  • Sunset. The memo ceases to have force two years after its issuance unless the OMB Director provides otherwise.

The executive order adds two duties of its own. Each LLM contract entered into after the OMB guidance must provide that decommissioning costs are charged to the vendor if the agency terminates for noncompliance after a reasonable period to cure. Each agency must adopt procedures to ensure procured LLMs comply within 90 days of the guidance [1].

What vendors are asked to provide

The memo’s appendix sets a minimum that every LLM solicitation must request [2]:

  • the acceptable use policy;
  • model, system or data cards, which the memo notes are seldom all available for one product;
  • end user resources such as tutorials and developer guides; and
  • a mechanism for end user feedback on outputs that violate the principles.

Depending on planned use, an agency may ask for more [2]. That can cover pre-training and post-training actions affecting factuality, system-level prompts, what content moderation filters restrict, and red teaming for bias. It can also cover training or development outside the United States and modifications made to comply with another government’s regulations. Beyond the model itself, the enhanced list includes bias evaluation results and methods, benchmark scores, enterprise controls such as configurable system instructions and source citation, and any controls added by a vendor that is not the original developer. The memo tells agencies to avoid, where practicable, requirements that compel disclosure of sensitive technical data such as model weights [2].

Three further points shape how this works in contracts. First, agencies must request this information when an LLM is built into another software product or service, as well as for standalone models [2]. Second, the memo recognizes that agencies often buy through resellers and integrators, and that what they can learn depends on the original developer’s willingness to cooperate through that channel [2]. Third, agencies should identify the requirements as material to eligibility and payment, to support termination for default where a vendor refuses corrective action [2].

For models an agency builds itself, the memo expects documentation of pre-training and training, evaluation, enterprise controls, and how the team handles feedback from end users [2]. For other generative tools, such as image, voice or multimodal generation, agencies should use the guidance to inform documentation requirements where practicable [2].

What this means for agency data and AI teams

In our view, the practical effect is a documentation trail for each model: the vendor’s cards and policies at award, a channel for users to report outputs, and the evidence an agency would need if it ever invoked the termination terms. The memo gives one example of fitting the principles to use: for an LLM that summarizes cybersecurity incidents for a security operations center, truth-seeking is likely more relevant than ideological neutrality [2]. The memo sits beside the governance duties in M-25-21, covered in our M-25-21 explainer, and our comparison of federal cloud and AI buying rules sets M-26-04 beside M-25-22 and other procurement rules. How it applies to a given contract is a question for each agency’s acquisition and legal staff.

Status as checked on November 14, 2026. This explainer describes the memo; it is not legal advice.

Moraph’s teams help organizations with work like this. See our solutions.

Spotted an error? Tell us and we will check it and correct the explainer where needed.

  • OMB
  • M-26-04
  • Executive Order 14319
  • large language models
  • AI procurement
  • Unbiased AI Principles
  • M-25-22
  • federal agencies