GSA Finalizes LLM Data-Safeguarding Clause for Federal AI Contracts

The General Services Administration has finalized the contract language that will govern how AI vendors handle government data inside large language models. On September 28, GSA issued clause 552.239-7001, “Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems,” as a GSA Regulation Deviation, the Government Contracts Legal Forum reported on October 6. It takes effect on October 19, though contracting officers may add it to procurements earlier.

A narrower clause than the drafts

The final text is the third version this year. A March draft drew heavy industry criticism, including for an outright ban on AI components “manufactured, developed, or controlled by non-U.S. entities.” A June revision dropped that ban. The final clause applies only when the government is specifically buying a system in which LLM functionality is a “material feature” and government data is submitted directly to or produced by the model. A contractor’s internal back-office use of LLMs, and products where AI is incidental, fall outside it.

Flow-down rules were rewritten as well. Instead of four defined supply-chain roles, obligations now track the task categories in the NIST AI Risk Management Framework and reach any subcontractor that handles government data. Requirements about origin or foreign control do not have to be flowed down to newly defined “Fully Open” models and open LLM components.

What changes for vendors

Contractors gain intellectual property protections. The government does not acquire ownership of their preexisting commercial products and tools, general capability gains that do not reveal government data are carved out of the IP assignment, and contractors keep their background data even when it is modified. The 72-hour incident notice now covers only incidents affecting an LLM used on the contract that may affect government data, and reports already filed under FedRAMP or CISA rules can satisfy it.

The June draft’s “Unbiased AI Principles,” which required a model to be a “neutral, nonpartisan tool” free of “ideological dogmas,” were replaced by a single duty to use reasonable efforts to prioritize accuracy, scientific inquiry and objectivity in factual answers. The government, however, can now suspend an LLM “at any time,” and contractors face decommissioning costs after a notice of noncompliance, capped at 25 percent of the affected order.

Why it matters

The federal government is a major AI buyer, and the clause sets baseline terms for GSA’s LLM purchases while the White House reshapes AI oversight; see our report on Jay Clayton’s appointment as AI czar and the voluntary self-policing accord signed with AI leaders. The clause still overrides vendors’ commercial agreements, a point the authors expect to keep drawing industry objections.

What to watch

Whether GSA issues revised companion clauses for the four supply-chain roles still mentioned in the text, how many contracting officers adopt the clause before October 19, and whether other agencies borrow its terms.