U.S. GSA Issues AI Procurement Rules: Contractors Barred From Training Models on Government Data

What Happened

The clause sits on page 129 of a memo signed by GSA Senior Procurement Executive Jeffrey Koses, titled "Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems." Three core rules: the government owns all government data (inputs and outputs included) plus custom developments; using government data to train, fine-tune or improve any LLM — including third-party-operated ones — is prohibited, as is using it for advertising, marketing, monetization or handing it to other parties; contractors must implement data-handling procedures including encrypted transmission and audit logging. The memo itself dates to January; the LLM section was signed off in July and takes effect October 19.

Two carve-outs narrow the scope: LLMs contractors use purely internally are untouched, and products where the LLM functionality is merely "incidental" are exempt. The government may run automated checks of a vendor's model for bias, truthfulness and safety and suspend its use at any time; if a contract is terminated over the vendor's fault, the contractor's decommissioning liability is capped at 25% of the affected order value. Quinn Anex-Ries, senior policy analyst at the Center for Democracy and Technology, publicly welcomed the final text as "a significant improvement" over the draft, with the "unbiased AI" compliance language "almost entirely" removed — industry had argued there was no testable way to comply.

Key Facts

1. Effective date: October 19; contracting officers may apply it now. It stays in force until GSA rescinds it or completes a formal rule with public comment.
2. Three red lines: No training/fine-tuning/improving LLMs on government data; no advertising, marketing or monetization use; no selling, licensing or out-of-scope retention of government data.
3. Contractor duties: Encrypted transmission, audit logging; security incidents reported within 72 hours; the government reserves automated-check rights over vendor models and may suspend use at any time.
4. Parallel move: Rep. Ro Khanna wrote to OpenAI, Anthropic, Google and others demanding disclosure of every known attempt by China or other hostile actors to steal model weights over the past two years, plus their defenses (per Reuters, via aipolicydaily 10/2/2026).

Context

GSA published a March draft and a June revision, both of which went through public comment, with industry objecting to the "unbiased" language and fuzzy definitions. The final version absorbed 75 comments, narrowed scope to "government-purchased AI," and dropped the ideological provisions. At the federal level, OMB memo M-25-22 already directs agencies to embed similar safeguards, and the FAR Council is working on government-wide rules.