TL;DR
GSA has issued a binding AI acquisition clause via class deviation, effective October 19, setting immediate compliance rules for any federal contract that delivers or provides government access to an AI system.
What happened
- GSA issued its AI acquisition clause as a class deviation from the GSA Acquisition Regulation, making it enforceable policy without waiting for formal rulemaking.
- Effective date is October 19, but contracting officers can apply it now and may modify existing contracts at their discretion.
- The clause followed over 75 public comments this summer, prompting significant rewrites from the June draft.
- Scope is narrowed: the clause applies only when the government buys AI, not when contractors use AI internally, unless those tools are delivered to or accessed by the government.
- Three categories of open AI are defined (fully open, open-weight, open-LLM components), with different foreign-control flowdown requirements for each.
Why it matters
- Every new federal AI contract is now subject to a uniform compliance framework, affecting the entire government technology vendor ecosystem immediately.
- The government reserves the right to suspend any LLM at any time, a stronger posture than the June draft, which required performance issues to be unresolved before suspension.
- Contractors must notify the government within seven calendar days of any material change that increases output bias, weakens safety guardrails, or degrades truthfulness, creating a live monitoring obligation.
- Liability for decommissioning costs after a for-cause termination is capped at 25% of the affected task or delivery order, a concrete financial boundary industry pushed for.
- Flowdown requirements now follow the data, not the model: subcontractors are only bound if they handle government data, reflecting Nvidia's publicly stated position from the comment period.
What to watch next
- Formal rulemaking timeline: GSA has not announced dates, and the deviation remains policy until it is either rescinded or codified through a public comment process.
- Government suspension actions: the expanded suspension right is untested; the first invocation will signal how aggressively agencies intend to police LLM outputs.
- Open-weight model treatment: the carve-out for fully open models but not open-weight models (weights only, no architecture or training data) could reshape how vendors structure their federal offerings.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.