Per-Claim Citations vs Per-Answer Citations

When an AI tool cites its sources, it usually cites them at the bottom. The answer runs for several paragraphs, and underneath sits a list of documents the system drew on. That is better than no citations at all, but it quietly leaves the hardest part of review to the reader: figuring out which specific claim came from which specific source. A list of five documents under a twelve-sentence answer tells you the answer is grounded somewhere. It does not tell you whether the one sentence you are worried about is grounded anywhere.

That gap is where unsupported claims survive. A reviewer reads a confident answer, sees a credible-looking source list, and moves on, without ever confirming that the specific figure or recommendation they are about to rely on is the part those sources actually support. Per-answer citations make the answer look sourced. Per-claim citations make each claim checkable, and the difference between those two is the difference between review that reassures and review that verifies.

Per-answer citation hides the weak link

The problem with citing at the answer level is that it averages over the claims. An answer might contain eight statements: seven of them well supported by the attached documents, and one that the model extrapolated, softened, or simply got wrong. A single source list at the bottom covers all eight equally, so the one weak statement is camouflaged by the seven strong ones. The reviewer has no signal pointing at the sentence that needs a second look.

This is exactly the kind of error that does damage, because it does not look like an error. The answer reads as a coherent whole, the sources are real, and nothing on the page distinguishes the claim that rests on solid evidence from the claim that rests on the model’s confidence. To catch it, the reviewer has to re-derive the link between every claim and every source by hand, which is the work the citations were supposed to save them.

Per-claim citation tells the reviewer where to look

Citing at the claim level inverts this. Instead of one source list for the whole answer, each individual claim carries its own link to the specific passage it came from. The reviewer can see, claim by claim, which statements trace back to the evidence and which do not. A claim with no supporting passage stands out immediately, because the absence is visible rather than buried in an average.

Qonera attaches citations at the claim level. Each claim in a synthesised answer carries the specific evidence passages it draws on, with clickable links straight to the source, and each claim is also marked with how well the evidence supports it. On the Multi Model path, the Conflict Heatmap tags each claim Green, Orange, Red, or Outlier based on whether the independent models agreed, so the reviewer sees both where the evidence is thin and where the models disagreed, at the level of the individual claim rather than the answer as a whole.

Verification becomes targeted, not exhaustive

The practical payoff is that review stops being all-or-nothing. Without per-claim citations, a careful reviewer has two options: verify everything, which is slow enough that it rarely happens, or spot-check and hope, which is fast but misses the camouflaged claim. Per-claim citations offer a third option: go straight to the claims the system has flagged as weakly supported or contested, verify those properly, and spend less time on the claims that are clearly grounded.

That is not about trusting the system to decide what is true. It is about the system pointing the reviewer’s attention at the right places, so the limited time a reviewer has goes to the claims most likely to be wrong. The human still makes the call. The citations just make sure the call is informed by where the evidence actually is, rather than by how confident the prose happens to sound.

The same principle sits behind incoming regulation

The same principle sits behind Article 13 of the EU AI Act, which expects high-risk AI systems to be transparent enough for the people deploying them to interpret and use the output appropriately. Interpreting output appropriately is hard when the link between a claim and its evidence is left implicit. Per-claim citation makes that link explicit, which is what transparency means in practice rather than in principle. Most of the obligations under the EU AI Act apply from August 2026, and teams that can already trace each claim to its source are close to what the transparency expectation pushes toward.

Citing the answer tells a reader the work was sourced somewhere. Citing each claim tells the reviewer exactly where, and exactly which claims to be careful about. For a quick personal answer the first is fine. For work that carries a firm’s name to a client, the second is what lets the team say not just that the work was grounded in evidence, but that they checked the specific places where it might not have been.

This article is for general information only and does not provide legal advice. Organisations should consult qualified legal counsel about how Article 13 and the EU AI Act apply to their specific systems, workflows, and obligations.