The General Services Administration (GSA) has issued an AI-specific acquisition memo restricting how contractors may use government data processed by large language models. Contracting officers can use its new clause now, although the stated effective date is Oct. 19, 2026. It is contract policy issued through a class deviation, not a final regulation.
What is GSA’s AI acquisition clause, and when can it be used?
The clause sets terms for covered GSA contracts involving large language models, or LLMs. An LLM generates responses to prompts; in a government contract, a prompt or source document supplied to the system may contain government data. The clause governs how contractors handle that data when their LLMs process it.
GSA used a class deviation to make the terms available to contracting officers without first completing formal rulemaking. According to FedScoop’s account of the issued policy, officers may use the clause before Oct. 19 and may choose to modify existing contracts. That does not mean every current GSA contract has been changed.
The policy follows several drafts. GSA said it released an initial version on Jan. 12, then published an updated proposal on June 17 and sought comments through Aug. 3. The agency said the revisions responded to feedback on the earlier draft. FedScoop reports that GSA received more than 75 comments over the summer; that figure does not show how many commenters supported or opposed particular provisions.
A separate White House budget office procurement memo directed agencies to address government data and intellectual-property rights when buying AI. Its guidance included contract terms barring the use of nonpublic agency data and outputs to train publicly or commercially available algorithms without agency consent. GSA’s clause sets out terms for LLMs in covered contracts.
Which uses of AI and government data does the clause cover?
The key distinction is whether an LLM processes government data as part of what a contractor provides to the government. GSA’s June proposal described inputs such as prompts, queries and source documents, as well as outputs that include generated content or changes to the inputs. A document sent for processing can therefore bring government data into an LLM system.
The reported policy does not cover every AI tool a contractor uses. It generally excludes a contractor’s internal business use of an LLM when the tool is not delivered to or accessed by the government, unless a contracting officer otherwise identifies that use as covered. It also excludes products whose LLM function is incidental to their main purpose.
Contractors may not use government data covered by the clause to train LLMs, inform advertising or sell it to third parties. Processing government information to fulfill a contract does not give a vendor permission to reuse it for those purposes.
A contract can involve a prime contractor and companies supplying parts of its service. The clause’s reach into that supply chain drew scrutiny during drafting. FedScoop reported that Nvidia chief external affairs officer Bruce Andrews argued obligations should depend on who handles government data, not who created the underlying model. The revised terms narrowed which subcontractors receive requirements.
What safeguards and reports must contractors provide?
The clause calls for encrypted transmission, which protects data while it travels between systems, and audit logging, which records activity for later review. It also imposes reporting and closeout duties. These are contractual requirements, not evidence that the safeguards have already prevented misuse.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Kevin Martin, a GSA program manager at Government Acquisitions Inc., described a 72-hour incident-reporting window and 30 days’ advance notice before a major model change, Nextgov/FCW reported. His account also identified a 120-day disclosure deadline. FedScoop reported a separate requirement to report certain material changes in output bias, safety controls, performance or truthfulness within seven calendar days.
For a vendor, compliance extends beyond protecting data when it first enters a system. Martin’s account also described deletion duties when a contract ends. The reporting does not quantify the cost of meeting these requirements. Martin told Nextgov/FCW that the remaining workload was substantial for products whose LLM features fall within the clause’s scope.
The terms changed in response to industry concerns, but the record does not establish that every requested change was accepted. Nextgov/FCW reported that some industry groups objected to an earlier provision on unbiased AI principles because, they said, compliance lacked a clear test. Their objection is not a finding that the government’s eventual assessments will or will not work.
Who checks the systems, and what happens after Oct. 19?
FedScoop reports that the government retains rights to assess covered LLMs, including checks of bias and truthfulness, and may suspend their use. The available reporting does not establish how often officials will use those powers or whether the checks will reliably detect problems.
For agencies and the public, the intended benefit is tighter control over government information sent to contractors’ LLMs. For vendors and subcontractors in scope, the trade-off is continuing data-handling and reporting obligations. Quinn Anex-Ries, a senior policy analyst at the Center for Democracy and Technology, welcomed the final terms as an improvement over earlier drafts, Nextgov/FCW reported. That assessment does not independently verify the safeguards’ effectiveness.
Oct. 19 is the clause’s stated effective date, not the start of a completed federal regulation. FedScoop reports that GSA could leave the class deviation in place, rescind it or pursue formal rulemaking; no schedule for that process was reported.
