Methods
Auditing an LLM answer against its sources
Split an answer into claims and inspect cited passages for support, inference and missing qualifications.
Problem
A fluent answer can combine supported statements, unsupported additions and citations that merely mention the topic. Source presence alone does not establish faithfulness. Auditing asks whether each material claim follows from the material available to the answer, with the same entity, time, scope and certainty. The audit must distinguish an omitted qualification from an outright contradiction. This method applies whether sources were retrieved automatically or supplied by a person.
Steps
Freeze the audit package
Keep the exact answer, question, cited source versions and passages available when the answer was produced. Record answer date and reader access scope. Later web content may differ from what the answer saw. If available context is incomplete, state that limitation and avoid claiming to have audited information that cannot be inspected.
Decompose material claims
Split the answer into assertions that can be checked separately, including numbers, dates, identity links, causal statements and negatives. Preserve qualifiers and implied comparisons where they affect meaning. A sentence can contain both a supported fact and unsupported inference; scoring only at sentence level hides that difference. Mark procedural or subjective text separately from factual claims.
Compare claim to passage
Locate exact evidence for each claim and ask whether it entails the claim under the same scope. Check entity, date, units, negation and whether the source reports a fact or an allegation. Mark direct support, support requiring explicit inference, contradiction or no support. A source discussing the same topic but not the asserted value provides no support.
Report and correct
Record each finding with a locator and explanation, then revise unsupported wording or declare uncertainty. Report the denominator of material claims, categories used and inaccessible evidence. If sources conflict, the corrected answer should surface the conflict rather than select one without grounds. Recheck the corrected answer against the same source versions before closing the audit.
Checks
- Every material claim maps to inspected evidence or an explicit gap.
- Entity, time, units and qualifications match.
- Inference and direct support are distinct.
- Answer and source versions are preserved.
Pitfalls
- Counting a topical citation as support.
- Checking only already-cited claims.
- Letting later revisions rewrite the original audit.
A worked example
A fictional assistant answers: "Halden Therapeutics AG's compound received orphan designation in 2024 and the trial enrolled 120 patients." The audit package holds the question, the answer, the three passages the assistant saw and the date. The answer is split into claims: orphan designation, the year 2024, a trial, an enrolment of 120.
The auditor finds the designation and the year in passage one. The trial appears in passage two, but the enrolment figure is 112, not 120, and the passage says planned, not enrolled. The third passage is about a different compound with a similar name. The findings: two claims supported, one contradicted in a detail, one with a qualification dropped, and a distractor passage that was not used. The corrected answer states 112 as planned enrolment, cites passage two, and the revised version is checked against the same passages. The numbers 2 of 4 supported are reported with the audit, not buried.
Variations and edge cases
Atomic claims matter. Work on factual precision for long-form text splits generated text into atomic facts and scores the share that are supported, which reveals partial errors that an overall judgement hides. For cited answers, benchmarks distinguish whether the answer is correct from whether the citations support it, since an answer can be right for reasons the citations do not show. Measure both.
Some claims cannot be audited. If the answer rests on material the auditor cannot see, say so. Subjective statements and advice are marked as outside the audit. Claims that need combining several passages are marked as inference, with the passages listed. Models may also produce statements consistent with the sources in tone and absent from them in fact, a failure that surveys of hallucination and studies of faithfulness in summarization describe at length.
Using automated judges
Automated judges can scale the audit, and they are fallible. Calibrate any judge against human audits on a sample, measure its agreement, and look at its errors by claim type. Do not let the same model family write and judge without a check. Keep humans on consequential answers and on disagreements between judges. Evaluation frameworks for retrieval-augmented systems provide metrics for faithfulness and context relevance, and these are best treated as screening tools.
How the method is checked
Plant known faults. Add to a test set answers with a wrong number, a dropped negation, a swapped entity and an unsupported addition, and confirm that the audit catches each. Re-audit a sample with a second auditor and compare. Verify that the corrected answers actually pass when rechecked against the same evidence.
References
- Min et al. (2023). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. arXiv:2305.14251.
- Rashkin et al. (2021). Measuring Attribution in Natural Language Generation Models. arXiv:2112.12870.
- Ji et al. (2022). Survey of Hallucination in Natural Language Generation. arXiv:2202.03629.
- Maynez et al. (2020). On Faithfulness and Factuality in Abstractive Summarization. arXiv:2005.00661.
- Es et al. (2023). Ragas: Automated Evaluation of Retrieval Augmented Generation. arXiv:2309.15217.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Test this method on one entity.
Name an entity where "Auditing an LLM answer against its sources" would be checked in practice, with its sources and its readers. We reply with what a first record would cover and what it would need from you.
Or write to [email protected] in your own words.