Model citations often come from memory not retrieved sources
Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation
Artificial Intelligence
Summary
Sometimes when AI systems generate answers with citations, the cited documents aren't actually the source of the information. The authors tested this by swapping true facts in the retrieved documents with false ones and checking if the model’s answers changed. They found that in many cases, the model relied on what it had memorized instead of the cited passages. Existing methods to verify citations often fail to detect this problem because the model picks citations that match its answer, whether or not they were actually used. The authors created a better way to detect true provenance but it is still not reliable enough to trust completely.
What this means in practice
- •For ai system developers: Improve tools detecting whether large language models actually use retrieved documents when generating answers with citations.
- •For fact-checking teams: Develop enhanced methods to assess the trustworthiness of AI-generated citations by focusing on model behavior changes with manipulated input documents.
Authors
Tanzeel Abbas, Tehseen Zia, David Windridge
Abstract
Retrieval-augmented generation systems attach citations to attribute claims to their sources. These citations frequently misstate provenance, and standard verification cannot detect this. We measure provenance by counterfactual value substitution: a fact in the retrieved passages is replaced with a false variant and the model re-queried. We find that across five sizes of one model family and a second comparator family, 1.2% to 15.7% of citable claims derive from parametric memory rather than the retrieved context. At the most conservative substitution distance these figures are 4.6% and 8.5% on the two models we tested at it, with the rate rising with model size within a family. We further find that natural language inference verification, the standard approach, gives results indistinguishable from chance across models, even when restricted to the cited passage. We determine that the cause is structural: a model answering from memory selects a citation that agrees with its output. Comparing ten attribution methods on identical claims, the contrast between the model's distribution with and without context best separates provenance. Acting on that signal does not improve provenance honesty at these base rates. The resulting detector is hence still not accurate enough to act on: it measures what the model knew rather than what it used, and separating those remains the open problem.