AS-2603_AI Use Transparency Requirements _SIGNED_04-23-2026.pdf

PRESCRIPTION: USE THIS CLAUSE IN ALL SOLICITATIONS AND CONTRACTS

ABOVE THE MICROPURCHASE THRESHOLD.

AS-2603 AI Use Transparency Requirements

Contractors must disclose all use of Artificial Intelligence (AI) in the products and/or services they deliver or use to deliver the products or services under the contract.

A list of AI Definitions can be found on the AI@HUD page on hud.gov.

1. High-Impact AI Disclosure (HUD):

The Government may determine, in the solicitation or during contract performance, the use or planned use of an AI System(s), AI Service(s), or AI technology (collectively referred to as AI System) meets the definition of High Impact AI, or there is a reasonable likelihood that such a use may arise during the life of the contract.

2. Vendor Transparency Obligations:

The Contractor shall not use any AI System in performing work under this contract without prior written authorization from the Government. Within ten (10) business days of award, the

Contractor must submit to the COR a request for written authorization for all planned use of

AI Systems the Contractor intends to use in performance of the contract. If the Contractor seeks to use a new or additional AI System, or discovers AI System after contract award, the

Contractor shall submit a request for written authorization to the COR at least ten (10) business days prior to implementation or discovery. The Contractor shall not use any AI

System(s) without prior written approval from the Government.

For each request for written authorization to use an AI System, the Contractor shall submit to the COR information to enable HUD to evaluate the proposed AI System(s). The Contractor shall make all AI System(s) documentation available in both technical and non-technical

(layperson) formats, tailored, as needed, to meet the requirements of the specific contract, and suitable for the Government’s internal review, oversight, and audit.

At a minimum, the Contractor shall submit the following:

A. Clear and complete documentation of the AI System(s) design, intended function, and operational context that includes, but is not limited to:

i. Documentation that demonstrates the AI System(s) complies with the

Unbiased AI Principles, including truth-seeking behavior and ideological neutrality

ii. System context diagram

iii. End-to-end data flow diagram

iv. Retrieval-augmented generation flow https://www.hud.gov/ai

v. Logical data model/schema diagram

vi. Indexing and refresh lifecycle diagram

vii. Access control and trust boundary diagram

viii. Sequence diagram for key use cases

ix. Observability and audit logging flow

x. Failure modes/fallback flow

xi. AI system name and version

xii. Hosting environment

xiii. Model type

xiv. Data sources

xv. Third-party dependencies

B. The purpose, function, and type of all AI Systems and technologies used in performing work under the contract.

C. A Software Bill of Materials (SBOM) for AI components and disclosure of third-party models.

D. Identify proposed AI Systems that would be used for automated decision making.

Provide a detailed description of human oversight and quality assurance of AI output.

E. Documentation outlining how the AI System(s) operates, including how outputs are generated, and how risks, including bias, privacy violations, inaccuracies, or lack of explainability, are mitigated.

F. A completed HUD AI Assurance Packet and written approval from the Contracting

Officer with concurrence of the HUD Office of the Chief Information Security

Officer (OCISO) or its designee.

G. Documentation outlining how the AI System is designed and tested to comply with the Unbiased AI Principles as outlined in Executive Order 14319, Preventing Woke

Al in the Federal Government (July 23, 2025):

Sec. 3. Unbiased AI Principles. It is the policy of the United States to promote the innovation and use of trustworthy AI. To advance that policy, agency heads shall, consistent with applicable law and in consideration of guidance issued pursuant to section 4 of this order, procure only those LLMs developed in accordance with the following two principles (Unbiased AI Principles):

(a) Truth-seeking. LLMs shall be truthful in responding to user prompts seeking factual information or analysis. LLMs shall prioritize historical accuracy, scientific inquiry, and objectivity, and shall acknowledge uncertainty where reliable information is incomplete or contradictory.

(b) Ideological Neutrality. LLMs shall be neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas such as DEI. Developers shall not intentionally encode partisan or ideological judgments into an LLM’s outputs unless those judgments are prompted by or otherwise readily accessible to the end user.

Consistent with the EO, Developers shall not intentionally encode partisan or ideological judgments into an AI System’s outputs unless those judgments are prompted by or otherwise readily accessible to the end user. OMB Memorandum M-

26-04, Increasing Public Trust in Artificial Intelligence Through Unbiased AI

Principles (December 11, 2025).

Unless specifically requested, the Contractor is not compelled to disclose sensitive technical data like specific model weights.

H. Documentation that supports the Unbiased AI Principles, including:

i. The Acceptable Use Policy. This is typically drafted by the original LLM developer to characterize and differentiate appropriate and inappropriate use of their product offering.

ii. Model, System, and /or Data Cards. All available information about the AI model, AI System, and/or data, including, summaries of the training process and details on the training sources, identified risks and mitigations, and AI model evaluation scores on AI models benchmarks.

iii. End user resources. May include product tutorials, developer guides, and tools that help end users ensure proper use of the AI System and maximize utility.

iv. Mechanism For End User Feedback. Describe mechanisms available to end users and the Government to report suspected violations of the Unbiased AI

Principles (e.g., hallucinations, factual inaccuracies, or ideological bias).

Provide the Contractor’s process and timelines for investigating, documenting, and remediating such violations.

I. Information sufficient for the Government to conduct an AI Impact Assessment, including risks related to fairness, explainability, privacy, and reliability.

J. A detailed description of how the AI was trained, how HUD data will be used to

“train” the AI System, and how HUD data will be labeled / designated as HUD data.

K. Confirmation of Contractor’s compliance with AS-2602, Privacy, IP, and

Government Data Use Requirements, and that HUD data will only be used for the product(s) and service(s) delivered to HUD under the contract.

L. Potential biases of training data and proposed mitigations for such biases.

M. Frequency of Contractor reviews of training data and AI System output(s) for potential bias or inaccuracies.

N. Detailed description of the Contractor’s method to technically prevent non-public data from being used in AI System training or refinement, including how those methods meet HUD requirements, which may exceed FedRAMP authorization requirements.

O. Information describing pre-training and post-training activities, including:

i. Actions undertaken that would impact the factuality and grounding of outputs;

ii. System-level prompts that provide natural language instructions to the AI model specifying guidelines on responding to user-generated prompts and queries particularly with regard to how an AI model responds when reliable information is incomplete, contradictory, or subject to individual interpretation;

iii. The type of outputs restricted via content moderation and safety filters;

iv. Use of red teaming as a means of continuously assessing the AI model to protect against incidents of bias in output;

v. Any AI System or model training or development that occurred outside of the

United States. Provide the type of activity and the country in which the activity and/or development occurred;

vi. Any modifications or configurations to the AI System or model undertaken to comply with any law or regulation from a government other than the U.S.

Federal Government.

P. AI model evaluation information, including:

i. Results of bias evaluations conducted by the Contractor and the methodology for performing such evaluations (e.g., testing prompt pairs for politically oriented topics);

ii. Benchmark scores for Contractor evaluations to measure an AI model’s bias, helpfulness, honesty, or accuracy when provided with ambiguous versus straightforward prompts. In some instances, benchmarks to compare performance across multiple languages will be necessary.

Q. Enterprise-level controls information including:

i. Governance tools, such as customizable system instructions that are additive to a base AI model’s system prompts or content generation filters;

ii. AI model evaluation tools that enable comparisons either between an AI model’s outputs, or across various AI models;

iii. Product features that require an AI model to cite sources of its outputs or otherwise provide visibility into the provenance of the AI model’s outputs.

R. Change management procedures for model updates.

S. Third party modification information including disclosure of additional controls to modify an output applied by the Contractor if the Contractor is not a direct developer of the AI System (e.g., classifiers, system prompts, fine-tuning, content moderation and filters).

T. Written confirmation that the Contractor will not:

i. Use publicly hosted AI System(s) to process HUD data, including, but not limited to CUI, PII, FTI, or classified data.

ii. Permit HUD data to be retained by third-party vendors for training.

iii. Introduce unapproved AI Systems and/or tools.

iv. Deploy AI Systems capable of autonomous decision-making without human validation, unless approved in writing by the Government.

U. Documentation to meet the requirements outlined in the AI Technical Requirements, which can be found on the AI@HUD page on hud.gov.

3. Monitor AI System Use and Performance.

The Contractor shall regularly monitor use and performance of each AI System and promptly rectify behavior defined as unacceptable. The Contractor shall ensure the AI System meets all use and performance standards of the contract, prior to deploying any new version. If a newly deployed version of an AI System fails to meet use and performance standards, the

Contractor shall promptly roll back or return the AI System to a prior approved version. The

Contractor shall submit a quarterly AI risk attestation and an annual bias reassessment in accordance with HUD requirements. The Government reserves the right to conduct reviews, including, but not limited to architecture, policy, security, and funding reviews, and may suspend or revoke authorization to use an AI System if risk exceeds acceptable thresholds.

4. Provide Notification of New AI System Features.

The Contractor shall notify and provide corresponding documentation to the COR for written approval to use an AI System, with details to the Government of new AI System enhancements, features, and/or components, at least 30 days prior to implementation of such https://www.hud.gov/ai enhancements, features, or components. Contractor shall not implement new AI System enhancements, features, and/or components without the Government’s written approval.

5. Ongoing Reporting:

The Contractor shall update the above documentation and disclosures as the AI System evolves, including if any new AI System, feature and/or component is added after the contract is awarded. As soon as practicable, but at least thirty (30) days prior to the deployment of major updates or newly integrated features, Contractor shall provide updated documentation and disclosures to the COR for approval. Contractor shall not implement major updates or integrated features without the Government’s written approval.

6. Subcontractor Compliance:

The Contractor shall flow down all AS-2603 AI Use Transparency requirements to its subcontractors that are providing AI components and/or AI Systems, under this contract and/or utilizing AI components and/or AI Systems in performance under this contract.

The Contractor shall provide an estimate of the cost to bring the AI System or AI Service into compliance with applicable Government requirements. Based on this information, the

Government will determine whether to continue, modify, or discontinue use of the AI System and/or AI Service.

The Contractor’s failure to adhere to the requirements of this Agency Specific clause could result in the Government exercising any rights available under the terms of the contract, including assessment of disincentives, withholding payment, liquidated damages, and/or actual damages, as appropriate. The Government reserves the right to stop work on this contract pursuant to this clause until the Contractor demonstrates compliance with this requirement.

Compliance with this Agency Specific clause is a material requirement for eligibility of contract award.

End Clause

X Francine L. Hemphill

Chief Procurement Officer