In short. A faithful citation links a supported statement to a document that causally influences its generation.

Definition: Citation faithfulness measures whether a cited document actually influences the statement an AI assistant generates. In the definition proposed by Wallat and colleagues, a faithful citation points to a document available during generation, supports the statement, and reflects a causal influence on it.

Key facts

Wallat and colleagues' Correctness is not Faithfulness in RAG Attributions, published in December 2024, separates citation support from actual source use.
The ALCE paper evaluates citation quality through recall, which checks support across statements, and precision, which identifies irrelevant citations.
ContextCite distinguishes sources that cause a model to generate a statement from sources that support its content.

How does faithfulness differ from correctness?

Citation correctness means that the referenced document supports the associated claim. Retrieval-augmented generation (RAG) supplies retrieved documents to a language model as context for producing an answer. Faithfulness concerns how the model uses that context.

In Wallat and colleagues' example, a model answers a question about Germany's capital and cites a document containing the answer. The same visible citation can result from reading the document or from recalling the answer from internal knowledge and finding matching support. The authors call the latter process post-rationalization.

Their formal definition requires support as well as causal influence. A document can therefore support a correct answer while the citation misrepresents where the model obtains its information. The answer text and link alone leave these processes indistinguishable in the example.

How does the paper test source use?

The December 2024 study tests Command-R+ on NaturalQuestions, a question-answering dataset. The system retrieves documents and generates answers with citations. The researchers take short cited statements from those answers and append them to other documents, then generate answers again.

One condition modifies documents the model previously considers relevant but leaves uncited. The paper's figure reports:

Relevant but previously uncited document condition
Modified question-answer pairs: 702
Original statement found again: 476
Modified document cited for that statement: 273
Reported rate among matched statements: 57%

The denominator matters: the percentage concerns cases where the original statement appears again after the context changes. The experiment tests whether short text matches draw citations toward modified documents. It supplies evidence of post-rationalization under this setup; the reported rate describes this experiment with Command-R+.

The authors identify an unresolved assumption: inserting irrelevant material must leave the model's underlying decision process unchanged for their test's interpretation to hold.

What can a website owner check?

A support audit compares an answer claim with the cited passage. The ALCE evaluation uses natural language inference, a task that assesses whether a premise supports a hypothesis, to automate this comparison. This measures the relationship between the document's content and the statement.

Testing influence requires intervening on the material supplied to the model. ContextCite calls removal of selected context a context ablation. It evaluates attribution by measuring how excluding sources changes the probability of the original response. This requires access to the generation setup and model probabilities.

For a website audit, the practical distinction is between verifying a linked page's support and experimentally tracing its influence. A record of a visible citation supports the former investigation; the causal tests described in these papers require additional access to the system.

FAQ

Can a correct citation still be unfaithful?

Yes. In Wallat and colleagues' example, the model can answer from internal knowledge and attach a supporting document afterward. The citation supports the claim while misrepresenting the document's role in generating it.

Does removing a source always settle the question?

ContextCite's limitations describe cases where several sources contain the same information: the response may change only when all are removed. The paper also explains that deleting a sentence can change the meaning of a remaining pronoun, complicating interpretation of the test.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer