DoD Unbiased AI Guidance for Contractors: 4 Documents to Prepare
Executive Summary
- The DoD unbiased AI guidance is an October 20, 2026 memorandum from Defense Pricing, Contracting, and Acquisition Policy that gives contracting officers three templates for applying Executive Order 14319 and OMB Memorandum M-26-04 to AI buys. The templates are optional and tailorable. The order and the OMB memo bind the agencies; your own obligations are whatever your solicitation, contract or subcontract actually says.
- For covered large language model buys, agencies must request four categories of transparency information: the developer’s acceptable use policy, a current model, system or data card, end user resources, and a feedback mechanism for outputs that may violate the principles. These are information categories, not necessarily four standalone files.
- When the standard contract language is incorporated, it makes compliance a material requirement for the whole period of performance, requires updated documentation after significant modifications, and sets a 15 business day deadline to investigate and answer a written notice of a problem output. That is a response deadline, not necessarily a deadline to finish the fix.
- Where incorporated, the language flows down to subcontracts at any tier for the acquisition, development or integration of LLMs, which can leave resellers and integrators promising documentation that largely sits with the model developer.
- Small contractors should assemble a four part documentation package now, read the actual terms of each solicitation, secure upstream cooperation, map who controls each model change, and rehearse a 15 business day output investigation before the first notice arrives.
In this article
- What Is the DoD Unbiased AI Guidance?
- Does the DoD Unbiased AI Guidance Apply to Your Company?
- What Documents Must Contractors Provide?
- What Happens After Award?
- Why Resellers and Subcontractors Have the Hardest Job
- How Does the DoD Guidance Compare With OMB, GSA and DOE?
- What About Existing Contracts and Option Years?
- How to Run a 15 Business Day Output Investigation
- What to Put in Teaming Agreements and Subcontracts
- A 30 Day Readiness Plan for Small Defense Contractors
1. What Is the DoD Unbiased AI Guidance?
The DoD unbiased AI guidance is a memorandum dated October 20, 2026 and signed by John M. Tenaglia, Principal Director of Defense Pricing, Contracting, and Acquisition Policy, that gives Department contracting and program staff standard templates for buying AI that meets the “truth seeking” and “ideological neutrality” principles. It is posted in the DPCAP Policy Vault as USA001389-26-DPCAP and runs nine pages including its attachments.
The DoD unbiased AI guidance implements Executive Order 14319 and OMB Memorandum M-26-04 inside the Department of War. It supplies three optional templates: a requiring activity memorandum, Section L proposal instructions and statement of work language. For covered large language model buys, agencies request four categories of transparency information, and the standard contract language, where used, requires continuing compliance for the period of performance.
The memo states the split plainly. Compliance with the executive order and the OMB memorandum “is mandatory,” while use of the attachments “is not mandatory,” and components may revise, tailor or supplement the language to fit the risk of each buy. For a contractor that means two things. You will not see identical wording in every solicitation, and the absence of these templates does not mean the underlying agency policy has gone away.
Keep one distinction in view throughout this article. The executive order and the OMB memorandum direct agency procurement actions. Contractors must follow the requirements actually included in their solicitations, contracts, orders and subcontracts. The DoD templates show the standard language you may encounter, but the exact text, including any tailoring, controls. An obligation that appears in a template but not in your contract is not automatically an enforceable term against you.
A word on names. The Department has used Department of War as a secondary title since Executive Order 14347 of September 5, 2025, and the memo itself is written on Office of the Under Secretary of War letterhead. Department of Defense remains the statutory name, and most searches and headlines still say DoD or Pentagon. This article uses DoD and DoW interchangeably for the same Department.
The memo sits on top of two earlier documents. Executive Order 14319, “Preventing Woke AI in the Federal Government,” was signed July 23, 2025 and published at 90 FR 35389. It directs agencies to procure only LLMs developed in line with two principles and to put compliance terms in each new LLM contract. OMB Memorandum M-26-04, issued December 11, 2025, turned that into procurement instructions and gave agencies until March 11, 2026 to update their policies. The DoD memo was issued about seven months after that deadline. That gap does not by itself show the Department took no earlier implementing action; we did not review component level policies.
Coverage has been steady. Federal News Network and DefenseScoop reported the memo between October 23 and October 26, and law firm alerts followed, including HWG’s November 2 advisory. On November 13, a Law360 report said the broad language could be applied unevenly by agency officials and could deter smaller, nontraditional contractors from competing. That concern is the reason this article exists.
2. Does the DoD Unbiased AI Guidance Apply to Your Company?
The DoD unbiased AI guidance reaches your company when a DoD solicitation or contract you are bidding on or performing carries language built from it, which typically happens when the requiring activity classifies the requirement as involving an LLM or, at its discretion, another type of AI. Internal use alone is not the test: the OMB memorandum excludes incidental administrative AI use that is neither directed nor necessary to fulfill contract requirements, not every use that happens inside your company.
The decision starts on the Government side. Attachment A is a memorandum the requiring activity signs and gives the contracting officer, who keeps a copy in the contract file. Its first block asks the requiring activity to pick one of four categories:
- An LLM, defined as in the executive order as a generative AI model trained on vast, diverse datasets that generates natural language responses to user prompts.
- A generative AI model other than an LLM, with examples ranging from vision language models to small language models.
- Discriminative AI used for a task such as classification, regression or object detection.
- No LLM or other AI at all.
If the first box is checked, the template says the four minimum transparency items “must be included in both the solicitation and the resulting contract,” meaning the solicitation should request the information and the resulting contract should identify the applicable documentation and performance obligations. If the second or third box is checked, including them is optional and left to the person signing the memorandum. That is a real difference for small firms. A computer vision or predictive maintenance vendor may see none of these requirements on one contract and all of them on the next, depending on who signed Attachment A.
Three carve outs come from M-26-04 rather than from the DoD memo. First, the OMB memo does not govern “AI used incidentally by a contractor during performance of a contract for administrative purposes,” such as AI you choose to use when it is not directed or necessary to fulfill requirements. Second, agencies are not required to apply it to LLMs acquired under a free, open source license, although they should perform due diligence on those models. That is discretionary treatment, not a prohibition, so an agency may still apply the requirements to an open source model. Third, it does not apply to national security systems as defined in 44 U.S.C. 3552(b)(6). Defense related work is not automatically a national security system under that definition.
Do not lean too hard on that third carve out. M-26-04 says applying the memo to national security systems is “encouraged” to the extent practicable, and the DoD memo does not repeat the exclusion. A defense program that touches a national security system can still carry this language if the component chooses to tailor it in. Read the actual solicitation.
Scale matters here. Deltek’s Clarity research, based on insights from more than 900 contractors, reported that as many as 90 percent of firms use AI in some capacity. That is a survey population, not a census, and it does not tell you how much of that use is incidental. The clearest cases are deliveries or integrations of models for Government use, but contractors should also assess other contract directed AI use rather than relying only on whether a model is a separately delivered item.
3. What Documents Must Contractors Provide Under the DoD Unbiased AI Guidance?
Under the DoD unbiased AI guidance, a solicitation for an LLM requirement that uses the standard language asks the offeror to submit four categories of documentation with its technical proposal: the LLM developer’s acceptable use policy, the current model, system or data card, a description of end user resources, and a description of the feedback mechanism the Government can use to report outputs that may violate the principles. Attachment A says this documentation “will be incorporated into the contract as a material requirement.” Your actual solicitation controls what you must submit.
These four items track Appendix A, Section 1.A of M-26-04 almost word for word. The OMB memo adds useful color. It notes that it is “uncommon” for a vendor to produce all three card types, so the memorandum does not require all three; the agency still assesses whether the card or cards supplied give enough information for the product. It also says the feedback mechanism “may be satisfied by a general inbox, specific product point of contact, or similar mechanism.” You do not need to build a portal. For readiness purposes, think of this as a four part documentation package, not four standalone PDFs.
Beyond the minimum, Attachment A offers an enhanced menu that the requiring activity can check item by item for higher risk buys. These enhanced items apply only when selected for a particular acquisition; they are not universal disclosures. Attachment B reserves this paragraph for higher risk requirements and tells the contracting officer to delete it entirely if nothing was selected.
| Tier | What the template lists | Who usually holds it |
|---|---|---|
| Minimum (required for LLMs) | Acceptable use policy; model, system or data card; end user resources; feedback mechanism | Mostly public documents from the model developer, plus your own feedback channel |
| Enhanced: pre and post training | Actions affecting factuality and grounding; system prompts; outputs restricted by moderation and safety filters; red teaming for bias; training outside the United States; changes made for a foreign government’s rules | Model developer; often confidential or only partly published |
| Enhanced: model evaluations | Results and methodology of bias evaluations; benchmark scores for bias, helpfulness, honesty or accuracy | Developer for base models; you, if you fine tune or wrap the model |
| Enhanced: enterprise controls | Governance tools such as custom system instructions or content filters; evaluation tools for comparing outputs; features that make the model cite sources | Platform vendor or integrator |
| Enhanced: third party modifications | Classifiers, system prompts, fine tuning, moderation and filters added by a vendor that is not the direct developer | You, if you are a reseller or integrator |
The last row is the one most small firms underestimate. If you add a system prompt, a retrieval layer that changes what the model says, a guardrail classifier or a fine tuned adapter, you are the “vendor who is not the direct developer,” and you are the party best placed to document the controls you added. Keep that record current, because it is also your best evidence when an output is challenged later.
Attachment B closes with a short sentence that carries most of the risk: “An offeror’s failure to provide the information required by this provision may result in the proposal being ineligible for award.” Missing required information may make a proposal ineligible rather than merely lowering its score, although the actual consequence depends on the solicitation and its evaluation rules.
There is one protection worth knowing. Section 4(a)(ii) of the executive order told OMB to avoid requiring disclosure of model weights or other sensitive technical data “where practicable,” and M-26-04 repeats that agencies should avoid requirements that compel disclosure of specific model weights. If a solicitation appears to ask for weights or source data, raise it during questions and answers rather than assuming you must comply or decline to bid.
4. What Happens After Award: Continuing Compliance and the 15 Business Day Response
When incorporated into the contract, Attachment C makes compliance with the two principles a material requirement for the entire period of performance, treats your proposal documentation as the compliance baseline, requires updated documentation after significant modifications, and requires a written investigation response within 15 business days of a contracting officer’s written notification. The instructions for Attachment C say the language should be incorporated “in its entirety” into the statement of work or performance work statement.
“Significant Modifications” is defined broadly: any change by the contractor or subcontractor “that could reasonably be expected to alter” the model’s adherence to the principles, including but not limited to changes to training data, model architecture, filtering and content moderation. Swapping to a newer model version, rewriting a system prompt or loosening a content filter can all qualify. Not every release is necessarily a significant modification. As an internal control, we suggest screening every release for whether it could reasonably affect adherence to the principles, and documenting the answer.
The notice process is specific. If an output “appears to violate” the principles, the contracting officer sends written notification. Within 15 business days you must investigate and respond in writing with your findings, whether there was a violation, the most likely cause, your proposed corrective actions and your proposed timeline. Fifteen business days is roughly three calendar weeks, excluding applicable holidays, and it includes the time to get answers from any upstream model provider. It is an investigation and response deadline. The template does not require complete technical remediation within those same 15 business days; the timeline for corrective action is something you propose.
GOVERNANCE INSIGHT
The 15 day clock is a records test before it is a technical test.
Exact reproduction of a nondeterministic model output may be impossible, but preserved prompts, outputs, model identifiers, retrieval material, configuration, logs and release records can support a defensible likely cause analysis. That requires knowing which model version, system prompt, retrieval sources and filter settings were live on the date of the output. Firms that already keep change records under ISO/IEC 42001 or the NIST AI RMF Manage function will answer faster. Firms that do not will spend most of the 15 days reconstructing history.
Materiality is the enforcement hook. Attachment A says the requirements must be identified as material “for determining contract eligibility and payment, and to support termination of the contract for default when the contractor refuses to take corrective action in identified cases of noncompliance.” In plain terms, refusing to fix an identified problem is framed as grounds for default termination, not just a performance rating issue.
One gap is worth noting. Section 4(b)(i) of the executive order directs agencies to include terms “providing that decommissioning costs shall be charged to the vendor” when a contract is terminated for noncompliance “following a reasonable period to cure.” The Attachment C template we read contains no decommissioning cost term. A component tailoring the language could add one, and GSA’s own clause does, so check every contract you sign rather than assuming the template governs.
Labelling compliance a material contract requirement also has a side effect lawyers will notice. Knowingly false claims or misleading certifications about these obligations could raise False Claims Act questions. But under the Supreme Court’s decision in Universal Health Services v. United States ex rel. Escobar, a contractual label of materiality or condition of payment is relevant but does not by itself establish materiality: falsity, knowledge, materiality and the other elements must still be proved. An isolated defective output or an ordinary performance dispute is not automatically a false claim. We identified no publicly reported enforcement action specifically under this October template as of November 16, 2026.
5. Why Resellers, Integrators and Subcontractors Face the Hardest Documentation Problem
Resellers, integrators and subcontractors face the hardest documentation problem because the template asks them to supply and maintain information about a model they did not build, while Attachment C, where incorporated, makes the contractor answerable for keeping that model compliant for the whole contract. The memo itself acknowledges the gap: the information available “will vary depending on the contractor’s role within the software supply chain,” with “more information generally being available from sources closer to the original LLM developer.”
M-26-04 goes further. It says it is “common” for agencies to access LLMs through resellers, intermediary deployers, software integrators and platform operators, and that in those cases the availability of product information depends on “the willingness of the actual AI developer to collaborate” through the distributor. That is OMB conceding that a small reseller may not be able to get enhanced documents at all.
The flowdown makes the problem concrete. Attachment C, paragraph (d), requires the contractor to insert “the substance of the mandatory requirements” in “all subcontracts at any tier for the acquisition, development, or integration of LLMs to be used or delivered in the performance of this contract.” A prime whose contract carries that language will push it to you. Whether you can push it further to a frontier model provider whose API terms are not negotiable is a different question.
Our reading, on which we found no published interpretation, is that a standard click through API agreement is unlikely to give you contractual leverage over an upstream developer’s model updates. Note also that the “Significant Modifications” definition covers changes made “by the contractor or subcontractor.” An unannounced change by the upstream developer may not fit that definition literally, but the duty to keep the model compliant “throughout the period of performance” still applies. How far that responsibility reaches depends on the actual obligation, any Government specifications, your provider relationship and your contract terms, which is why it is worth negotiating information and change notice arrangements before award.
There are practical ways to narrow that exposure. Pin to a specific model version where the provider allows it. Subscribe to the developer’s deprecation and change notices. Keep your own evaluation set of factual and politically sensitive prompts and run it after every upstream release. Our analysis of what frontier providers actually pass down to their API customers found the same pattern in Europe: documentation rights exist on paper, but timing and depth depend on the provider.
6. How Does the DoD Unbiased AI Guidance Compare With OMB, GSA and DOE Rules?
The DoD unbiased AI guidance is distinctive for its fixed 15 business day response period, which the OMB memo does not specify. The Department of Energy also imposes extensive lifecycle, remediation, reporting and subcontract requirements, while GSA uses a different clause structure that dropped the itemized principles from its final text.
| Source | Date | Scope | Key contractor duty |
|---|---|---|---|
| EO 14319 | Signed July 23, 2025 | LLMs procured by federal agencies | Compliance terms in each new LLM contract; decommissioning costs charged to the vendor on termination for noncompliance after a reasonable cure period |
| OMB M-26-04 | December 11, 2025; sunsets after two years unless extended | All agencies; excludes national security systems and incidental administrative use; application to free open source LLMs not required but due diligence expected | Four minimum documents; optional enhanced documents; requirements identified as material |
| DoD DPCAP memo | October 20, 2026 | DoD components; templates optional and tailorable | Four documents mandatory for LLM buys; updates after significant modifications; 15 business day response; any tier flowdown |
| GSAR 552.239-7001 (deviation) | Effective October 19, 2026 | GSA solicitations and contracts with covered LLM products and services | Reasonable efforts to prioritize accuracy and objectivity; the decommissioning cost remedy is capped at 25 percent of the affected order |
| DOE AL 2026-05 | May 8, 2026 | DOE and NNSA buys where AI is a predominant function; M&O contracts and national security systems excluded | AI impact assessment before solicitation (or before award for simplified acquisitions); clause with advance notice of material changes, annual reverification, corrective action at no additional cost, withholding and all tier flowdown |
The GSA comparison matters for firms that sell through the Schedule as well as directly to DoD components. Our analysis of the GSA AI clause walks through the issued GSAR text, which replaced the June draft’s list of principles with a single reasonable efforts accuracy duty and capped decommissioning costs after a termination for noncompliance with the clause. The cap applies to that decommissioning cost remedy, not to all contractor liability or all Government remedies. The RGO PDF is a compilation that GSA updates over time, so check the current version before relying on any clause text. The DoD template keeps the two principles by name and attaches a fixed clock, so the same product may face different obligations depending on the buying channel.
The DOE Acquisition Letter 2026-05 is a useful contrast because it front loads work onto the Government and then adds a detailed clause. The requiring office categorizes impact as low, moderate or high and completes an AI impact assessment covering bias testing, explainability and human oversight before issuing a solicitation, or before award for simplified acquisitions. Its “Ensuring Unbiased AI Principles” clause then requires advance written notice of material model changes, annual reverification, corrective measures at no additional cost, a cure period backed by withholding and other remedies, and flowdown to all subcontract tiers. It excludes management and operating contracts for now.
For background on how M-25-22 and the broader 2026 procurement stack fit together, see our earlier guide to the 2026 federal AI procurement rules. The short version is that the policy layer is shared, while the clause text, clocks and remedies differ by agency.
7. What About Existing Contracts and Option Years?
Existing DoD contracts for LLMs can pick up these requirements by modification, and OMB identifies the next extending option exercise as the intended latest point for practicable modifications. That is not a statement that every existing contract automatically changes at that time. M-26-04 says agencies “should, to the extent practicable, modify existing contracts for LLMs” and that the modification “should occur, at the latest, prior to exercise of any option that extends the period of performance.”
The executive order uses similar language in Section 4(b)(ii): agencies should revise existing LLM contracts “to the extent practicable and consistent with contract terms.” That phrase matters. A bilateral modification needs your agreement, and if new documentation or response duties add cost, you are entitled to raise consideration. Before agreeing to added obligations, assess the modification authority, scope, consideration and pricing consequences, and do not sign an option modification that adds Attachment C language without pricing the work it creates.
Timing is also bounded at the other end. M-26-04 states that it “shall cease to have any force or effect two years after the date of its issuance,” which is December 11, 2027, unless the OMB Director provides otherwise. Contract terms already signed by then remain in your contract unless the contract or a later modification provides otherwise, so treat the sunset as a policy date, not a release from obligations you agreed to.
If your firm is also working through cybersecurity obligations on the same programs, the two tracks share evidence. Our piece on AI tools, CUI and the CMMC Phase II suspension covers the data handling side that the unbiased AI template does not touch.
Getting your AI governance in order
Dynamic Comply helps small and medium sized businesses build a real AI governance foundation using the NIST AI RMF and ISO/IEC 42001.
8. How to Run a 15 Business Day Output Investigation
A 15 business day output investigation should follow a fixed sequence: preserve the evidence, reproduce the output, isolate the layer that caused it, decide whether it is a violation, and propose a fix with dates. Attachment C names five things your written response must cover, so build the process around those five headings from day one.
We suggest splitting the clock roughly like this, as a planning assumption rather than anything the template requires:
- Days 1 to 2. Log the notice, identify the date and context of the output, and freeze copies of the prompt, response, model version, system prompt and retrieval sources live at the time.
- Days 3 to 6. Try to reproduce the output on the same configuration and on the current one, and record whether testing reproduced the behavior, how often, what configuration differences exist and what uncertainty remains. A failed reproduction is not proof the reported output never occurred. Ask the upstream provider for release notes covering the period, sharing prompts and outputs only as classification, CUI, privacy, security and contract restrictions allow.
- Days 7 to 10. Isolate the cause: base model behavior, your system prompt, retrieved content, a filter or user prompting. The executive order’s second principle expressly allows ideological content that is “prompted by or otherwise readily accessible to the end user,” so the user’s own prompt is a legitimate finding.
- Days 11 to 13. Draft findings, a violation determination with reasons, corrective actions and a timeline. Have someone outside the engineering team review it.
- Days 14 to 15. Final review and delivery to the contracting officer, with a record of transmission.
Two points deserve care. First, the response must state “whether there was a violation.” You are allowed to conclude there was not, but the conclusion has to rest on evidence, and the truth seeking principle includes acknowledging uncertainty where reliable information is incomplete or contradictory. Second, your proposed timeline may become an agreed corrective action commitment. Distinguish proposals, dependencies and agreed milestones, promise only what you control, and say which steps depend on an upstream provider.
The NIST AI RMF is a practical scaffold here. Its Govern function covers roles and accountability, Measure covers your evaluation set, and Manage covers incident response and change control. NIST AI 100-1 is free and maps cleanly onto the five headings Attachment C requires. Following it does not create a presumption of compliance, but it gives your response a structure a contracting officer will recognize.
9. What to Put in Teaming Agreements and Subcontracts
Teaming agreements and subcontracts should allocate three things before award: who supplies each of the four documents, who must give notice of model changes and how fast, and who does what inside the 15 business day window. If the prime’s contract carries Attachment C, its flowdown paragraph requires the prime to pass the substance to covered subcontracts whether or not your teaming agreement anticipated it.
As a subcontractor, ask for these terms:
- A clear statement of which LLM, version and configuration you are responsible for, so you are not answerable for a layer the prime controls.
- An internal deadline that leaves you real time. If the prime receives a notice on day 1 and gives it to you on day 6, you have nine business days, not fifteen.
- Cooperation and information rights running both ways, including access to the output logs you will need to reproduce a problem.
- Allocation of any decommissioning or reprocurement costs that a tailored prime contract adds.
As a prime, ask your subcontractors and model suppliers to represent that the documents they gave you are current, to notify you of significant modifications in advance where they can, and to support investigations within a set number of business days. Where the upstream provider will not sign anything, document that fact and your compensating controls, such as version pinning and your own evaluation set. That documentation supports your explanation to a contracting officer; it does not guarantee acceptance or excuse an unmet contract term.
10. A 30 Day Readiness Plan for Small Defense Contractors
A small defense contractor can be ready for the DoD unbiased AI guidance in about 30 days by assembling the four document file, documenting its own model controls, setting a change log, and rehearsing one investigation. This plan is a practical baseline, not a safe harbor, and a tailored solicitation can ask for more.
- Week 1: inventory. List every product or service you sell to DoD that contains an LLM or other AI model, the model provider, version and how the customer accesses it. Separate internal administrative AI use, which the OMB memo excludes.
- Week 1: assemble the four part package. For each model, save the current acceptable use policy, model or system card, end user resources and your feedback channel, with retrieval dates.
- Week 2: document your own layer. Write down every system prompt, filter, classifier, retrieval source and fine tuning you apply. This is the third party modifications disclosure, and only you can produce it.
- Week 2: build an evaluation set. Assemble 50 to 100 factual, ambiguous and politically sensitive prompts relevant to your use case and record baseline outputs. This is our suggested size, not a template requirement.
- Week 3: change control. Define what counts as a significant modification for your product and require a documentation check before each one ships.
- Week 3: contracts. Read the actual unbiased AI terms in each active solicitation, contract and modification, and review teaming agreements and active subcontracts for flowdown, notice and cost allocation terms.
- Week 4: tabletop. Run one mock 15 business day investigation on a real output and time each step.
The same file supplies reusable evidence for three audiences, with agency specific supplements: the DoD contracting officer, GSA if you sell through the Schedule, and any commercial customer asking the questions buyers increasingly ask. That is also where a management system such as ISO/IEC 42001 earns its keep, because it turns a one time scramble into a routine that survives staff turnover.
Talk to Ross about your AI governance needs
Every business has different AI governance requirements. Let us talk about yours.
Frequently Asked Questions
What is the DoD unbiased AI guidance?
It is an October 20, 2026 memorandum from Defense Pricing, Contracting, and Acquisition Policy, signed by John M. Tenaglia, that gives DoD contracting staff three templates for applying Executive Order 14319 and OMB Memorandum M-26-04 to AI purchases. The templates are optional and can be tailored, while the order and the OMB memo are mandatory for covered agency acquisitions. A contractor’s own obligations are the terms actually included in its solicitation, contract or subcontract.
Does the woke AI executive order apply to contractors who use ChatGPT internally?
Generally not, if the use is incidental. OMB Memorandum M-26-04 says it does not govern AI used incidentally by a contractor for administrative purposes when that use is not directed or necessary to fulfill contract requirements. Internal use that is directed or necessary to perform the contract is not automatically excluded, and the clearest cases are models delivered or integrated for Government use.
What documents does DoD require from LLM vendors?
For covered LLM buys, agencies must request four categories of information: the developer’s acceptable use policy, the current model, system or data card, a description of end user resources, and a mechanism for reporting outputs that may violate the principles. These are information categories, not necessarily four separate files. Higher risk buys can add system prompts, bias evaluations, red teaming and foreign training details where the agency selects them.
How long do contractors have to respond to a noncompliant AI output notice?
Where the Attachment C template is in the contract, 15 business days after the contracting officer’s written notification. The response must cover findings, whether there was a violation, the most likely cause, proposed corrective actions and a proposed timeline. It is a response deadline, not necessarily a deadline to complete the fix.
Do the unbiased AI requirements flow down to subcontractors?
Yes, where the template is used. Attachment C requires the contractor to insert the substance of the mandatory requirements in all subcontracts at any tier for the acquisition, development or integration of LLMs used or delivered under the contract. Resellers and integrators should expect primes to pass this language to them.
What happens if an AI vendor will not fix a noncompliant model?
Where incorporated, the template treats compliance as a material requirement intended to support termination for default when a contractor refuses corrective action. Executive Order 14319 also directs agencies to charge decommissioning costs to the vendor after a reasonable cure period, although the DoD Attachment C template does not itself include that term.
About the author
Ross J. is the founder of Dynamic Comply, an AI governance, compliance, and cybersecurity consulting firm based in Leesburg, Virginia. He brings more than 15 years of federal cybersecurity experience across the Department of State, the Department of Defense, and the Department of Homeland Security, and holds the CGRC certification along with credentials as a GSDC AI Compliance Lead Implementer and Auditor and Certified Ethical Hacker.
Related articles
- GSA’s AI Contract Clause Takes Effect October 19. Internal and Incidental AI Use Can Fall Outside It. Small Business Status Does Not.
- 90% of Government Contractors Use AI. CMMC Phase II Is Suspended. Your CUI Obligations Are Not.
- Federal AI Procurement Rules 2026: A Contractor’s Guide
- OpenAI Reported One of Three Agent Incidents to Brussels. If Your SaaS Product Runs on a Frontier API, Here Is What Actually Flows Down to You.
This article is provided for general informational purposes and reflects the state of the law as of November 2026. It is not legal advice. Regulations in this area are changing quickly. Confirm current requirements and consult qualified counsel before making decisions for your organization.