Does Executive Order 14319 regulate private companies using AI?

The order directs federal agency procurement of LLMs. Its stated purpose distinguishes federal acquisition from regulation of model functionality in the private marketplace. A private company may still encounter the requirements as a vendor or reseller, or as an integrator or platform operator selling a covered LLM to a federal agency. Other statutes and regulations, plus contracts and sector rules, can create separate duties for private organizations. Executive Order 14319 should be recorded as the source for this federal procurement policy rather than treated as a universal AI-content rule.

Which AI systems fall within M-26-04?

The memorandum applies to LLMs procured by covered agencies, including models integrated into other software products or services. Deployment method and later modification leave that core scope intact, as does intended agency use. Agencies must also consider listed factors when deciding whether to apply the requirements to internally developed language models or procured generative systems for image and voice, as well as multimodal output. National security systems are excluded. Incidental administrative AI use by a contractor also falls outside the memo when the agency neither directs nor requires that use.

What was required by March 11, 2026?

Agencies had to update procurement policies and procedures so covered LLM contracts include requirements addressing the Unbiased AI Principles. The updated procedures also had to provide a process for agency users to report outputs that violate those principles. The deadline concerned agency policy and process. Each solicitation and contract, each model and evaluation, and each reported output and corrective action still needs its own dated record.

Can a vendor benchmark prove compliance?

A vendor benchmark supports the claim measured by its documented prompt set and method, its scoring process and model version, and its language and test date. It remains one component of the procurement record. Agency acceptance work should test the model in the intended use, with prompts and configurations relevant to that use. Recurring evaluations become appropriate when a provider changes the model or an intermediary changes a control. They also become appropriate when reported outputs reveal a new failure pattern. Vendor evidence and agency observations should remain separately labeled.

When does the OMB memorandum expire?

M-26-04 states that it will cease to have force or effect two years after its December 11, 2025 issuance unless the OMB Director provides otherwise. Agencies and vendors should monitor OMB for an extension or replacement, as well as any later implementation instruction, rather than treating the original memo as permanent.

Preventing Woke AI Executive Order: A Federal Procurement Evidence Map

Executive Order 14319 sets two principles for large language models procured by federal agencies: truth-seeking and ideological neutrality. The White House issued the order on July 23, 2025. OMB followed on December 11 with M-26-04, which translates those principles into solicitation terms and vendor-document requests, along with agency procedures and a March 11, 2026 policy deadline.

The Preventing Woke AI executive order creates a federal procurement test. Agencies need evidence about how a vendor develops and evaluates an LLM, then configures and supports it. They also need a separate record of how the procured model behaves in the agency's intended use.

TL;DR

Executive Order 14319 applies its Unbiased AI Principles to federal LLM procurement.
OMB M-26-04 requires contractual terms for covered solicitations issued after December 11, 2025.
Agencies had until March 11, 2026 to update procurement procedures and create a channel for reporting outputs that violate the principles.
Vendor disclosures describe development and control claims. Routed output evaluations record observable behavior under defined test conditions.
National security systems sit outside the memorandum's requirements, and the memo sunsets two years after issuance unless OMB extends it.

Executive Order 14319 establishes the procurement policy

The July 23, 2025 executive order directs agency heads to procure LLMs developed in accordance with two Unbiased AI Principles.

The first principle is truth-seeking. Under the order, an LLM should prioritize historical accuracy and scientific inquiry, with objectivity, when responding to requests for factual information or analysis. It should acknowledge uncertainty when reliable information is incomplete or contradictory.

The second principle is ideological neutrality. The order describes LLMs as neutral and nonpartisan tools, and directs developers to avoid intentionally encoding partisan or ideological judgments into outputs unless the user prompts for those judgments or can readily access them.

These are procurement requirements for federal agencies. The order itself says the federal government should hesitate to regulate model functionality in the private marketplace. That boundary separates this action from a general private-sector AI rule. The broader America's AI Action Plan contains many federal policy signals, while Executive Order 14319 creates a specific acquisition instruction for covered LLMs.

The order also directs agencies to place compliance terms in new LLM contracts after OMB issues its guidance. Those terms must address vendor-funded decommissioning costs when an agency terminates a contract for noncompliance following a reasonable opportunity to cure.

M-26-04 defines the operational scope

OMB Memorandum M-26-04, dated December 11, 2025, applies to LLMs procured by covered executive agencies regardless of deployment method or later modification, as well as agency use. The scope reaches a standalone model and an LLM embedded in another software product. It also reaches an LLM delivered through an integrator or platform operator.

The memo excludes national security systems as defined in 44 U.S.C. 3552(b)(6), although it encourages application where practicable. It also excludes incidental contractor use for administrative purposes when the agency neither directs that use nor requires it for performance.

Free and open-source LLMs receive separate treatment. The memo says agencies are not required to apply its requirements to models acquired under a free, open-source license, while recommending due diligence against the Unbiased AI Principles and other federal AI policy before use.

Solicitations or orders issued after December 11, 2025 must include contractual requirements addressing the principles. Existing LLM contracts should be modified where practicable, at the latest before an option extends the period of performance. Agencies also had until March 11, 2026 to update procurement policies and establish processes through which users can report outputs that violate the principles.

The memo loses force two years after issuance unless the OMB Director provides otherwise. A control register should therefore carry the issuance date and sunset condition, plus the current status and any extension or replacement.

Vendor evidence begins with four required disclosures

M-26-04 establishes a minimum transparency threshold for LLM solicitations. Agencies must request four categories of vendor material:

Acceptable use policy: The original developer's description of appropriate and prohibited product use.
Model, system, or data cards: Available summaries of training and identified risks, along with mitigations and evaluation scores.
End-user resources: Product guidance and developer documentation, plus tutorials or other material supporting appropriate use.
Feedback mechanism: An inbox, product contact, or similar route for reporting outputs that conflict with the principles.

This packet is procurement evidence. It tells an agency what the vendor claims, what documentation exists, and where responsibility sits across the software supply chain. It should preserve the document version and provider, the model family and applicable service, and the delivery date and contract reference.

A glossy model card with no service identifier is weak evidence. I would reject any procurement binder that puts four green ticks beside these requirements while the actual model route remains unnamed. Picture the folder on the conference-room table: four divider tabs and a printed benchmark chart, with a red sticky note asking which production endpoint the chart covers. That sticky note carries more operational value than the cover page.

The AI governance guide for procurement leads provides the surrounding workflow for supplier commitments and model changes, plus incident duties and exit evidence.

Additional transparency connects claims to model behavior

An agency may request more information when the planned use requires stronger assurance. M-26-04 lists several enhanced transparency areas.

Pre-training and post-training disclosures can cover factuality and grounding work, along with system-level instructions and restricted output categories. They can also cover safety filters and red-team activity, development outside the United States, and modifications made for foreign regulation.

Model-evaluation disclosures can include the methodology and results of bias tests. OMB gives paired prompts on politically oriented topics as one example. The memo also names benchmark results for bias and helpfulness, plus honesty and accuracy, across ambiguous and straightforward questions, with multilingual comparisons where relevant.

Enterprise controls form another evidence category. These can include configurable system instructions and output filters, tools for comparing models or responses, and features that provide citations or other provenance signals. Intermediaries should also disclose the classifiers and prompts they add to the underlying model, along with their fine-tuning, moderation, and filters.

The memo advises agencies to seek enough information for assessment while avoiding requests for model weights or other sensitive technical data where practicable. That balance belongs in the solicitation. Procurement should specify the claim it needs to assess and the artifact that supports the claim. It should also specify the service or model configuration to which the evidence applies.

Output evaluations require a separate record

Vendor evidence and observable output evaluations answer different questions.

A model card can describe a vendor's test methodology. A system-prompt disclosure can explain intended behavior. A benchmark result can summarize performance over a defined dataset. Each artifact remains evidence about the vendor's development process or assessment.

An agency-run evaluation records what a named model produced when given a defined prompt under a particular configuration. The test record should preserve:

The approved model route and provider.
The model or deployment identifier available at test time.
The prompt set and evaluation version.
The intended agency use and applicable principle.
Relevant system instructions and configurable controls.
The raw response or a controlled reference to it.
Human or automated scoring criteria.
Reviewer identity and disposition, with a timestamp.
Retest triggers for model and prompt changes, plus filter or route changes.

Paired prompts can reveal inconsistent treatment across equivalent factual requests. Ambiguous and straightforward prompts can test uncertainty handling. Multilingual sets can expose behavior hidden by an English-only sample. The LLM evaluation tools guide covers the mechanics of building repeatable test sets and preserving comparable results.

A clean evaluation result supports acceptance under those test conditions. It offers no proof about undisclosed training decisions or future model versions. It also offers no proof about traffic sent through another route. Procurement retains the vendor packet, while engineering or an evaluation owner preserves observed outputs and test conditions. Joining the two records through stable model and contract identifiers creates a usable evidence chain.

Contract terms should make the evidence enforceable

M-26-04 tells agencies to identify relevant transparency requirements as material to eligibility and payment. That designation supports termination for default when a vendor refuses corrective action after identified noncompliance.

A covered contract should name the required disclosures and update cadence, along with the feedback route and corrective-action process. It should define which model or service versions the commitments cover. Material changes need a notice rule, especially changes to the foundation model and system instructions, filters and evaluation methods, or the hosting arrangement and intermediary controls.

The contract record should also connect reported output concerns to a response process. Agency procedures can capture the output and request context, then classify the alleged principle violation. They can preserve the applicable configuration and notify the vendor, then record the cure decision. Procurement owns the contractual consequence. Technical reviewers assess the reproduced behavior. The agency's designated official owns the final compliance determination.

This structure follows the same discipline as AI model inventory management: every route needs a stable identity and owner, plus status and change history. A product name on an invoice leaves too much room for model substitution.

DeepInspect

DeepInspect can preserve the observable request-and-response portion of this evidence for authenticated HTTP traffic deliberately routed between agency users or agents and LLM endpoints. A routed evaluation can carry the caller and destination, the policy version and timestamp, and the decision outcome. The resulting record can link a test prompt and controlled response reference to the model route used at that moment.

Vendor training records and hidden development choices retain their existing owners, as do contract materiality and national security classification. Final compliance determinations also retain their existing owners. DeepInspect also leaves direct and off-gateway traffic outside its evidence boundary. The defensible design connects procurement disclosures to bounded output evaluations without pretending that one record proves the other.

Book a demo today.