TL;DR
The GSA has issued an immediately effective AI acquisition clause governing all new federal contracts that procure AI, setting enforceable standards for LLM accuracy, bias testing, and supply-chain vetting before formal rulemaking is complete.
What happened
- GSA issued a class deviation from its Acquisition Regulation on Oct. 2, 2026, making the AI procurement clause effective Oct. 19 but usable by contracting officers right now.
- The clause applies only when the government buys AI, not when contractors use AI tools internally, unless those tools are delivered to or accessed by the government.
- After more than 75 public comments this summer, GSA dropped contested language around "unbiased AI principles" and bans on "partisan or ideological judgments," replacing it with a "reasonable efforts" standard for factual accuracy.
- The government reserves the right to suspend any LLM at any time and to run automated assessments for bias, truthfulness, safety, and unsolicited ideological content.
- Contractor liability for decommissioning costs after a for-cause termination is capped at 25% of the affected task or delivery order.
Why it matters
- This is de facto federal AI procurement law right now: as a class deviation, it governs contracts until GSA rescinds or codifies it through formal rulemaking, which has no set timeline.
- Flowdown requirements now follow the data, not the model's authorship: subcontractors only face obligations if they handle government data, a direct win for Nvidia and other industry commenters who pushed this framing.
- Open-weight models (weights only, no architecture or training data) do not receive the foreign-control exemption granted to fully open models, creating a meaningful compliance split in the open-source AI market.
- Prime contractors must use "best efforts" to vet AI subcontractors, a stricter standard than the prior "due diligence" language, raising the bar for supply-chain accountability across the federal vendor ecosystem.
- A seven-calendar-day mandatory disclosure window for material increases in output bias, safety degradation, or truthfulness decline puts real-time compliance pressure on vendors with live government deployments.
What to watch next
- Whether GSA posts formal rulemaking dates: the deviation is indefinite until codified, so the timeline for a permanent rule will signal how much the current text can still shift.
- How contracting officers apply the suspension right: the removal of the prior "until performance issues are satisfactorily addressed" qualifier gives the government broader, open-ended authority to pull LLMs from contracts.
- Market response from open-weight model providers: the carve-out gap between fully open and open-weight models could push vendors to publish more training data and architecture details to qualify for the foreign-control exemption.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.