Citations and Source Attribution in AI Interfaces

In 2023 four commercial generative search engines were audited sentence by sentence. Only 51.5% of the generated sentences were fully supported by the sources they cited, and only 74.5% of the citations actually supported the sentence they were attached to (Liu, Zhang & Liang, 2023, Evaluating Verifiability in Generative Search Engines, Findings of EMNLP, arXiv:2304.09848). The superscript numbers looked identical whether they were right or wrong. That is the design problem: a citation is presented as evidence, but it is itself model output, with its own error rate, rendered in a visual language that borrows the authority of scholarship.

Designing for probabilistic output treats citations as the highest-leverage verification affordance. This concept asks what happens when that affordance is itself unreliable, and when users ignore it.

What a citation claims, and how to measure it

A citation marker attached to a statement makes two distinct claims. The support claim says the cited passage entails the statement. The provenance claim says the statement was produced from that passage. Interfaces conflate them, and systems break them independently. A post-hoc attribution pipeline such as RARR generates text first, then searches for evidence and revises unsupported content (Gao et al., 2023, RARR: Researching and Revising What Language Models Say, Using Language Models, ACL, arXiv:2210.08726). Its citations can satisfy the support claim while the provenance claim is false by construction. For a user deciding whether to believe a sentence, support is what matters; for an auditor asking why the model said it, provenance is.

Support is measurable. Let \(S\) be the set of generated statements that warrant verification and, for each \(s \in S\), let \(C_s\) be the citations attached to it. Following Liu et al.,

Recall asks whether the claims are backed. Precision asks whether the markers are honest. A system can raise precision by citing less and raise recall by citing everything, which is why the two are reported together. Automatic versions replace human judgement of entailment with an NLI model; the ALCE benchmark built on that idea found that even the best systems it tested lacked complete citation support about half the time on ELI5 questions (Gao, Yen, Yu & Chen, 2023, Enabling Large Language Models to Generate Text with Citations, EMNLP, arXiv:2305.14627).

A small calculation shows why per-sentence rates understate the user's problem. If each sentence is independently supported with probability \(0.515\), the chance that a six-sentence answer is supported throughout is \(0.515^6 \approx 0.019\). Errors cluster on hard queries, so independence is an idealisation, but the direction holds, and the reader cannot tell which sentence failed without checking all of them.

The trust a marker buys

The most uncomfortable result in this area is that citations raise trust independent of their accuracy. In a live experiment, answers carrying citations were trusted more than answers without them, and this held when the citations were random; participants who actually opened and checked the citations reported lower trust (Ding et al., 2025, Citations and Trust in LLM Generated Responses, AAAI, arXiv:2501.01303). The marker functions as a credibility cue processed heuristically, not as evidence evaluated analytically.

Liu et al. found a related tension at the system level: across engines, citation precision was strongly inversely correlated with perceived utility (\(r = -0.96\)). Their explanation is that engines which paraphrase or copy closely from sources are easy to verify and feel less useful, while engines that synthesise freely read as more helpful and cite less faithfully. The quality users reward and the quality that makes the answer checkable pull in opposite directions.

Design moves that respect the error rate

Cite at the grain of the claim. A single source list at the bottom of an answer gives recall no place to fail visibly. Sentence- or clause-level anchors let a reader check the claim they care about without auditing the whole answer.

Show the passage, not the page. Linking a long document turns a quick check into a search task. Highlighting the supporting span in a hover or side panel is what makes verification cheaper than the original task, which is the condition the verification-cost problem says an assistant must meet.

Render unsupported content differently. If the system knows a sentence has no citation, or its entailment score is low, saying so is more honest than a uniform paragraph. The design debate here is real: some practitioners argue that visible uncertainty markers clutter answers and train users to ignore them, others that uniform presentation is exactly what produced the trust effect above.

Distinguish retrieval from synthesis. A sentence that joins two sources into a conclusion neither states is not supported by either citation alone, and marking it with both implies otherwise.

When it breaks

Support is judged against a snapshot. Pages change or disappear; a citation that entailed the claim at generation time may not when the user clicks, and the interface rarely knows.

Entailment is itself contested. Human raters disagree on partial support, and automatic NLI judges inherit their own errors, so a "92% citation precision" figure carries the uncertainty of whoever judged it.

The source can be wrong. Perfect attribution to a low-quality page is a faithful citation of a false claim. Attribution metrics measure the link, not the truth at the end of it.

Nobody clicks. Every move above lowers the cost of checking; none creates the motivation to check. When stakes are low, the rational user skims, and the citation's main effect is the trust it lends.