# Government Buyers Specify Grounded AI Systems

A procurement commentary referencing an August 2025 paper argues that government AI contracts should specify system architecture and behavior rather than rely only on a headline hallucination or error-rate metric. It highlights retrieval-based grounding and raises whether systems should be considered successful when they abstain because the available record does not support an answer. No agency, solicitation, funding, or procurement timeline is identified.

- Solicitation writers can translate this approach into explicit requirements for retrieval-based grounding and system behavior when answers are unsupported by the record.
- Evaluation plans can distinguish unsupported answers from appropriate abstentions, rather than relying on a single error-rate measure.
- AI contractors should be prepared to describe how their systems retrieve and ground information and how they handle questions the source record cannot answer.

**Jurisdictions:** federal
**Industries:** Information Technology
**Topics:** Artificial Intelligence
**Published:** October 14, 2026

### Sources
- [The paper that "proved hallucinations can never be fixed" is from August 2025 and names the mitigation itself: retrieval. A government AI contract should specify an architecture, not an error rate. Should "not answerable from the record" count as a pass?

https://t.co/QKQmNIeGHc](https://x.com/miguelamigot/status/2110461187942940797) - twitter-fed-procurement