GSA’s New AI Clause Is Now in Effect: What AI Vendors and Investors Need to Know

Key Takeaways:

• The U.S. General Services Administration’s (GSA) new AI clause gives the government ownership of fine-tuned model weights and other custom developments built from government data under GSA large language model contracts. Vendors can’t use that data to train models for other customers without written authorization, and they must delete what they built from it when the contract ends.
• AI vendors should negotiate rights before award, document their Background Data and pre-existing IP, and consider architectures that keep government data out of their model weights. Investors and acquirers should confirm in diligence which models and datasets the company actually owns.

As of October 19, 2026, the government owns the fine-tuning work an AI vendor does on government data under General Services Administration (GSA) contracts, and the vendor has to delete it when the contract ends.

That is the practical effect of GSAR 552.239-7001, “Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems.” GSA finalized the clause on September 28, 2026, as a deviation to the GSA Acquisition Regulation, and it took effect as of October 19. It is among the first contract terms used across the government to treat fine-tuned model weights as government property.

For companies that sell AI to the government, this is not a compliance footnote. It decides which parts of your model you can carry to your next customer, and which parts you must rebuild. In other words, it is a question about what your company is worth. For investors and acquirers, it also changes what a target company owns.

What the clause does

The clause applies to GSA contracts, orders, and Blanket Purchase Agreements (BPA) for large language model (LLM) systems, where LLM functionality is a material feature and government data is submitted to or produced by the LLM. The clause will impact any agency buying through GSA, including the Department of War (DoW) and civilian agency orders placed on GSA vehicles. Its core terms fall into five groups:

• Ownership. The government owns its data and all “Custom Developments,” which include modifications from training or fine-tuning. Any rights the contractor obtains in improvements that incorporate or derive from government data are assigned to the government automatically. The contractor keeps its base model, its Background Data, and pre-existing tools used across customers. General capability gains that don’t derive from government data are carved out.
• Prohibited uses. Absent express authorization in writing by the contracting officer, the contractor may not use government data to train or improve models for other customers. It also may not use that data for advertising, marketing, sales, or business strategy, retain it beyond the contract, or sell or license it. Feedback belongs to the government, whether it is generated by government personnel, the contractor, or automated processes, which likely includes reinforcement-learning signals.
• Deletion. At the contract end, the contractor must delete fine-tuned weights, embeddings, indexes, and caches built from government data, and document the deletion to the contracting officer in writing.
• Override. The clause controls over the contractor’s End-User License Agreement (EULA), terms of service, and click-wrap terms. The clause does not override applicable law, regulation, or mandatory government contract clauses.
• Flowdown. The obligations flow to subcontractors and model providers that handle government data.

While the rule underwent several changes before being finalized, three things never changed: the training ban, government ownership of what is built from its data, and deletion at contract end. Those appear to be core policy positions. Contractors may want to consider planning around them rather than spend negotiating capital fighting them.

What the clause does not cover

Models that aren’t LLMs. The clause defines an LLM as a generative model that produces natural-language responses. So vision-language models and other models that answer in text are potentially covered. However, a perception model, an autonomy stack, and a signal classifier are outside the clause. Those deliveries fall under the traditional data-rights framework. There, the contractor generally keeps title and the government’s license depends on who paid for the development.

Direct Department of War (DoW) contracts. The clause applies only to GSA contracts and orders under them. DoW orders placed through GSA schedules or BPAs are covered, but DoW’s own contracts and OTAs are not. DoW has no equivalent yet on ownership or deletion of fine-tuned weights, but that does not leave DoW contractors free. Controlled Unclassified Information (CUI) may be used only for a lawful government purpose, and information derived from it stays controlled. A model fine-tuned on CUI is likely to be treated the same way. Moreover, DoW contracting officers may borrow GSA’s language now that it exists in writing.

Congress is also acting. The version of the 2027 NDAA passed by the House (H.R. 8800, § 861) would make government purpose rights the default for all technical data and computer software delivered to DoW, unless the contractor proves entitlement to more restrictive rights by clear and convincing evidence of private development funding. This rule would apply to model deliverables whether or not they are LLMs. The Senate bill has a similar provision.

Three takeaways for AI vendors

1. Reuse the base model, rebuild the fine-tune. Plan your business model around the line the clause draws. What you bring to the contract stays yours. What you build from government data stays with that contract and is deleted at the end. Reuse with another agency, or within DoW, requires the contracting officer’s written authorization. If the data includes CUI, you will also need sign-off from the agency that designated it, which is often a different office.
2. Negotiate before award, not after. The contracting officer’s written authorization is now the only lever for broader rights. Before you sign, ask for a written definition of the general improvements you keep, a deliverable defined as the fine-tuned adapter rather than your base model, and an inventory of your Background Data. Treat that inventory as an IP audit. Document what you developed independently of government data and consider patent filings on training and pipeline methods before they touch government data, so the carve out for general capability gains is documented. And review your model providers’ terms now, since the clause flows down to them.
3. Treat architecture as a legal decision. The clause attaches to what is built from government data. Retrieval-based designs that keep government data out of your weights keep the ownership question away from your weights. Choose your architecture with your counsel in the room.

A diligence flag for investors and acquirers

In diligence on a company that sells AI through GSA, ask which models and datasets are Background Data, which are Custom Developments subject to government ownership and deletion, and whether any contracting officer authorizations for reuse exist.