The General Services Administration (GSA), the purchasing agency of the U.S. federal administration, has published a clause regulating the purchase of large language models (LLM) and AI agents. Its text notably prohibits using State data to train these systems. According to Nextgov, the services were to apply it no later than October 19, 2026.
Clause 552.239-7001, “Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems,” is included in the new part 552 of the GSA purchasing regulation. It is published in memorandum RGO-2026-01, signed on January 15, 2026, by Jeffrey Koses, head of GSA procurement, then supplemented throughout the year. According to the firm Crowell & Moring, the GSA published it on September 28, 2026. As of October 22, its online version counts 224 pages.
An exemption pending the regulation
This is a “class deviation”: the GSA modifies its regulation by exemption, as part of the overhaul of federal purchasing rules launched by decree 14275. A formal regulatory procedure, including public consultation, must follow. The clause applies to calls for tenders issued from its effective date, and the contracting officer decides the fate of open consultations and already awarded contracts.
An initial draft, linked to MAS Refresh 31, had been criticized by the industry. It excluded components manufactured, developed, or controlled by non-American entities, according to Crowell & Moring. A revised version was then submitted for comments in the Federal Register of June 17, 2026. It merely prefers LLMs from an entity incorporated in the United States.
State data: no training nor resale
The obligations described below appear in the June draft, whose essentials for the final version are confirmed by Nextgov and Crowell & Moring. State data, whether input or output, cannot be used to train, refine, or improve an LLM, including third-party ones, nor to develop a model for other clients. They cannot feed the supplier’s advertising or commercial decisions, and resale is prohibited.
The provider only has a license limited to contract execution, technical support, and de-identified security telemetry, unless otherwise authorized in writing. At the end of the contract, it must permanently erase the data and specific developments, including the weights of fine-tuned models, embeddings, indexes, and caches, including backups, and report this in writing. Two 72-hour deadlines are added: one to report a substantial breach, the other to notify a security incident, followed by a daily update until resolution.
The definition of LLM retained is broad. It includes the model, but also hosting, system prompts, knowledge bases, retrieval-augmented generation (RAG) mechanisms that allow the model to rely on external documents, APIs, safeguards, and workflows.
American preference, exceptions, and ongoing contracts
The clause does not cite any foreign text, but it targets the laws, regulations, or policies of a “non-U.S. federal government.” A European publisher may remain admitted, provided it announces 30 days in advance any adaptation of its model to the rules of a non-American government that substantially changes its responses. According to the source, the application of the term “federal” to Union law remains to be clarified. Moreover, no core component may be developed or operated by an entity under the influence of a foreign “adversary” government, as defined by regulation 15 CFR 791.4.
Two exceptions render the clause ineffective. The first concerns an LLM limited to the provider’s internal systems, not delivered to or accessible by the State. The second targets a product whose AI is accessory, not required by the contract, not used by the administration, and without processing its data. A chatbot intended for users, however, falls within the scope.
The scope remains that of purchases made by the GSA. Claude from Anthropic, Gemini from Google, and ChatGPT from OpenAI were offered to federal agencies through the Multiple Award Schedule in August 2025. These ongoing contracts do not automatically fall under the new clause. In case of termination for cause after an ineffective formal notice, the supplier must pay dismantling costs capped at 25% of the order value, according to Crowell & Moring. The June draft left this percentage to be fixed contract by contract.
