Dateline: October 2, 2026 — Washington, D.C.
The General Services Administration has issued a binding AI acquisition clause for federal contracts, issued as a class deviation from its regulations and effective October 19. Why it matters: it’s the first federal procurement rulebook written specifically for AI — and any vendor selling AI to the government will have to live by it.
What happened
According to an agency regulation-overhaul memo updated Monday, GSA will impose its AI acquisition clause as a class deviation from the GSA Acquisition Regulation — making it usable immediately, with formal effectiveness October 19. A class deviation means it operates as policy until GSA rescinds it or codifies it through formal rulemaking, which will include a public comment period. The final terms follow two rounds of public comment — over 75 comments this summer plus pushback from industry groups — and incorporate several changes vendors asked for.
The key provisions vendors need to know
The clause only applies when the government is buying AI. It doesn’t cover AI tools contractors use internally, as long as those tools aren’t delivered to or accessed by the government — though it does now expressly cover AI-enabled document and productivity tools. Standout requirements:
- No training on government data: contractors can’t use government data processed by their LLMs to train models, inform advertising, or sell to third parties, and must use encrypted transmission and audit logging.
- Factfulness over ideology tests: the controversial “unbiased AI principles” language is mostly gone. Contractors must instead use “reasonable efforts” to make LLMs respond factually, prioritize accuracy and objectivity, and acknowledge uncertainty.
- Government audit rights: the government reserves the right to run automated assessments of the LLM for bias, truthfulness, safety, and unsolicited ideological content — and can suspend use of the model at any time.
- Seven-day change reporting: contractors must report any material change that increases output bias, decreases safety guardrails, or degrades performance or truthfulness within seven calendar days.
- Open-model tiers: fully open models (architecture, weights, code, and data published) are exempt from foreign-control requirements — but open-weight models that only publish weights don’t get the same exception.
- Capped liability: contractor liability for decommissioning costs after a for-cause termination is capped at 25% of the affected task or delivery order, and flowdown requirements now reach only subcontractors that handle government data.
Why it matters
Federal procurement terms have a way of becoming industry terms — what’s written into a GSA clause today shows up in enterprise contracts tomorrow. This clause is notable for what it doesn’t do as much as what it does: GSA listened to industry and dropped the untestable “unbiased AI” mandate, narrowed scope so internal tools are excluded, and — in a line worth quoting — made obligations follow the data, not the model’s authorship. The open-model distinction is the sleeper detail: it gives fully open models a compliance advantage in federal bids while pointedly not extending the same to open-weight-only releases. Tested Intelligence take: if you’re a vendor with any federal pipeline, the seven-day change-reporting clock and the government’s right to benchmark your model for “truthfulness” are the two provisions to operationalize first — they have teeth, and suspension is at the government’s discretion.
FAQ
When does the GSA AI clause take effect?
October 19, though contracting officers can use it now, and existing contracts can be modified at the officer’s discretion.
Does it apply to AI a contractor uses internally?
No — only when the government is buying AI. Internal-use tools that aren’t delivered to or accessed by the government are excluded, unless the contracting officer says otherwise.
Can contractors train models on government data?
No. Contractors may not use government data processed in their LLMs to train models, inform advertising, or sell to third parties.
What happened to the “unbiased AI principles” requirement?
It was removed after industry pushback over how compliance could be tested. The clause now requires “reasonable efforts” toward factual, accurate, objective responses — while the government keeps the right to assess models for bias and ideological content itself.
Sources: FedScoop; Nextgov/FCW