Source attribution
Overview
Source attribution (also AI citation) is the mechanism by which an AI system associates statements in a generated answer with the sources that support them, typically presented as inline links or numbered references. Attribution is the basis of citation rate and a prerequisite for users to verify AI answers.
Source attribution is widely reported to be unreliable: studies and audits have found high rates of citations that do not support the associated claim, or that point to non-existent sources, with reported error rates spanning a wide range depending on system and methodology.[1][2] Because the concept itself is contested and unstandardized, "attribution" can mean anything from a relevant link to a verified claim-to-source mapping.
How it works
Attribution is produced in different ways depending on architecture:
- Retrieval-grounded systems attach citations to the documents retrieved for the answer; accuracy depends on whether the cited document actually supports the generated claim (see Faithfulness vs Groundedness).
- Post-hoc attribution generates citations after the answer, which can produce plausible but unsupported references.
Evaluating attribution requires checking, per claim, whether the cited source exists and supports the statement — distinct from merely counting that a citation is present.
| Term | Refers to |
|---|---|
| Source attribution | Linking answer claims to underlying sources |
| Grounding | Anchoring an answer in sources at generation time |
| Groundedness | Whether each claim is supported by the provided context |
| Hallucinated citation | An attribution to a source that does not exist or does not support the claim |
Source attribution is not the same as correctness: a present, well-formatted citation may still fail to support the claim it accompanies.
Examples
- An answer footnotes a study URL that, when opened, does not contain the cited statistic — a citation present but unsupported.
- A model invents a plausible-looking paper title and author that do not exist — a hallucinated citation.
See also
References
- ↑ LLM Pulse. "Source Attribution in AI: what it is and how to improve it." https://llmpulse.ai/blog/glossary/source-attribution-in-ai/
- ↑ Research on fabricated references in LLM outputs documents large-scale generation of non-existent citations. See e.g. studies of hallucinated citations in model outputs.
