On July 14, the Coalition for Common Sense in Government Procurement (Coalition) participated in a public listening session concerning the General Services Administration’s (GSA) proposed Artificial Intelligence (AI) clause for the Multiple Award Schedule (MAS) program. Here is the statement, based on initial member input, that the Coalition provided at the listening session:
The Coalition is a nonprofit association of over 340 small, medium and large businesses selling commercial goods and services to the Federal Government, including through the MAS program. We have members that create, utilize, integrate and manage AI, as well as members that represent and advise clients on the interpretation of contract clauses.
The Coalition submitted extensive comments with respect to the initial draft AI clause and we commend GSA for accepting and considering industry input and deciding to take more time and receive additional input on this important, complex, cutting-edge topic. Collaboration between industry and Government is the best way to arrive at a solution that meets the needs of the Government without potentially depriving the Government of the best that the commercial market has to offer.
We would like to mention the following aspects of the clause that we believe require further attention:
We support GSA’s effort in paragraph (a) to limit applicability of the clause by stating that the clause does not apply when the Large Language Model (LLM) functionality is “incidental” to the primary purpose of the core requirement being procured or to LLMs embedded in common commercial products. However, the term “incidental” is undefined and subjective. We suggest further clarifying that LLM functionality is “incidental” when the contract does not require delivery of an AI-enabled capability and no Government-specified or Government-furnished AI tool is mandated as part of contract performance. Without such clarification, contracting officers could potentially insert the clause in contracts that involve the routine professional use of commercial AI tools, which is not GSA’s intent. The phrase “embedded in a common commercial product” also needs further clarification. The term “common commercial product” is undefined, and the examples given are limited to consumer tools. We suggest further clarification to confirm that this extends to enterprise software platforms with integrated AI capability. In addition, the term “embedded in” is unreasonably narrow. In practice, LLM functionality is integrated into commercial products through multiple technical architectures, including direct embedding, Application Programming Interface (API) calls, and third-party integrations. Thus, we suggest the phrase “used with or integrated into” instead of embedded. Finally, guidance should emphasize that the clause might not apply to every order under a contract, or to every discrete segregable task or line item under a contract.
In paragraph (d) we recommend amending the requirement to report “any” non-adherence to the clause. The requirement is overly broad and burdensome and is inconsistent with other clauses where contractors are not expected to report any non-compliance but instead report on mandatory disclosure requirement issues or cyber security incidents. An overly broad reporting requirement will result in overreporting trivial matters, which will waste industry and government time, increase costs, and distract the Government from focusing on meaningful, material issues. The clause already contains a 72-hour notification requirement for cyber incidents and a separate notice requirement where the LLM has been modified or configured to comply with non-US statutes, regulations, or policies. These and existing contractor disclosure requirements are sufficient to protect the Government’s interests.
With respect to the liability for decommissioning in paragraph (j), allowing the contracting officer to select a percentage of contract value is arbitrary and unnecessary. The clause already provides that the contractor is liable for reasonable decommissioning costs and there are existing well-established mechanisms for the government to recover costs associated with terminations for default. This potential for the imposition of costs that far exceed actual damages increases risk and will increase costs for the Government and ultimately American taxpayers.
The definition of “Data Outputs,” the assertion of full ownership over all Government Data including Data Outputs derived from Background Data, and the automatic assignment provision operate to transfer the contractor’s intellectual property rights in any output incorporating pre-existing Background Data. These provisions pose serious risks to contractor intellectual property and are inconsistent with the current data rights framework. Contractors should retain all pre-existing intellectual property rights including Background Data incorporated into or reflected in Data Outputs. The Government’s ownership interest should extend only to the novel content generated by the LLM or Government-originated content that does not derive from the contractor’s Background Data or pre-existing intellectual property.
The Coalition will be submitting written comments to the proposed rule on behalf of its members. We commend GSA for listening to industry feedback and thank GSA for providing us with the opportunity to speak today.