Provide citations, source tracebacks, or reasoning chains for generated content.
Produces the chain from output back to source: which passages informed which statements, which prompt segments drove which behaviour, what intermediate reasoning was involved. The difference between a system that asserts and one that shows its working.
Regulated decisions, audit contexts, and anywhere a person must defend the output to a third party rather than merely accept it.
Low-stakes assistance where attribution overhead exceeds the value of traceability.
Attribution must be verifiable, not decorative. A citation the reader cannot follow to a real, retrievable source is worse than none, because it manufactures confidence.
Post-hoc rationalisation presented as attribution. A model asked to explain its answer will generate a plausible explanation whether or not it reflects how the answer was produced — and this is genuinely difficult to distinguish from real feature attribution.
An auditor following a machine-generated explanation of a balance movement back through the specific journal entries and match records that produced it, and finding each one.