GSA's AI Procurement Rule Takes Effect Monday — Without Waiting for a Final Rule
Starting Monday, October 19, companies that sell artificial intelligence tools to the U.S. government have to comply with the federal government's first binding, enforceable AI-specific contract rule — and it didn't come from Congress, the White House, or even a completed federal regulation. It came from the General Services Administration's contracting office, which pushed a new data-safeguarding clause into force through a procurement mechanism called a "class deviation," months before the agency's own formal rulemaking on the subject is finished.
The clause, numbered 552.239-7001 and titled "Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems," sits on page 129 of a broader acquisition-overhaul memo signed by GSA Senior Procurement Executive Jeffrey Koses. GSA floated an early draft in January, circulated a revised version for public comment over the summer, and then — rather than wait for that rulemaking to finish — issued the language as an immediately usable deviation from the GSA Acquisition Regulation. Contracting officers can write it into new solicitations now and can amend existing contracts at their discretion. A formal rule, with its own comment period, is still coming; GSA has not said when.
A Rule With Teeth, If Narrower Teeth
The clause applies only when an agency is actually buying AI and large-language-model functionality is a material part of what it's buying — not when a contractor merely uses an LLM internally, and not when AI is "incidental to" a product's main purpose, such as spell-check inside word-processing software. Within that narrowed scope, the obligations are specific: contractors cannot use government data to train or fine-tune their models, to inform advertising, or to resell to third parties. Embeddings and fine-tuned weights built on government data must be deleted at contract closeout. Vendors must report security incidents within 72 hours and must notify the government within seven calendar days of any change that measurably increases output bias, weakens safety guardrails, or degrades accuracy — on top of separate 30- and 15-day notice-and-access windows before swapping in a major or minor model version. The government reserves the right to run its own benchmarks on a deployed model for bias and truthfulness at any time, and to suspend its use outright.
That's a retreat from GSA's original ambition. The January and June drafts would have required contractors to certify adherence to "unbiased AI principles" and avoid "partisan or ideological judgments" — language industry commenters said had no workable compliance test. The final version drops nearly all of it in favor of a "reasonable efforts" standard to keep outputs factual. GSA also capped a contractor's liability for decommissioning costs after a for-cause termination at 25% of the affected task or delivery order, and narrowed which subcontractors the clause flows down to — only those that actually design, develop, deploy, or operate the model, not every vendor in a supply chain that happens to touch it.
Industry Split, Watchdogs Cautiously Pleased
The retreat didn't stop Palantir from trying to kill the rule outright. In August comments, the company's outside counsel argued GSA lacks independent legal authority to finalize the clause, that it would violate the Federal Acquisition Streamlining Act by imposing non-standard commercial terms, and that — under the "major questions doctrine" — regulating something of this economic significance requires a clear delegation from Congress that doesn't exist. Palantir warned the rule would simply push agencies toward contract vehicles outside GSA's control, with its lawyers writing that the clause would likely keep the company from offering federal customers the solutions they need through GSA contracts.
Menaka Kalaskar, head of U.S. government legal and contracting, Palantir: "If the clause is really needed, then it can be included in the [Federal Acquisition Regulation overhaul]."
Most of the roughly 79 contractors and trade groups that filed comments by the August 3 deadline took a less confrontational line, asking GSA to narrow the clause's scope and sharpen vague terms rather than scrap it — and the final text reflects a lot of what they asked for. Quinn Anex-Ries, a senior policy analyst at the Center for Democracy and Technology, which had pushed for stronger safeguards, called the finished clause "a significant improvement from prior draft versions." A GSA spokesperson said only that the agency had been "highly collaborative throughout this rulemaking process" and was "thoughtfully integrating stakeholder feedback."
The episode is a reminder of where binding AI policy is actually getting made in Washington right now. With multiple federal AI-accountability bills stuck without a path to a floor vote and the White House's own frontier-AI task force still only months into a 120-day review, GSA's procurement authority — the government's power to set terms on what it buys, worth hundreds of billions of dollars a year — has become one of the few levers that has produced an enforceable deadline rather than another pledge, probe, or subpoena. It didn't require a single vote.