A plausible answer is not the same as sound evidence.

AI-generated answers often include citations that look authoritative: a link to a journal, a government domain, or a well-known publisher. When those links are followed, the page is missing, the domain has changed, or the anchor never existed.

The problem is not just broken links. It is the gap between the confidence of the answer and the fragility of the evidence. A user who clicks a citation assumes the source was checked. If the source cannot be found, the claim becomes unverifiable.

This failure is common after model training data ages, after publishers restructure URLs, or when retrieval systems hallucinate plausible-looking references.

A citation can be real while still being wrong. The source exists, the URL resolves, but the page does not contain the claim it is supposed to support. The AI system has either misread the source, overgeneralised from it, or attached the wrong reference to a generated sentence.

This is one of the hardest failures to spot by eye, because every individual element looks reasonable. The citation is from a credible domain. The claim sounds plausible. Only by comparing the two does the mismatch become clear.

For high-stakes answers — medical, legal, financial, public-policy — this kind of error can turn a helpful summary into a misleading one.

The web rewards repetition. A claim made once can be syndicated, quoted, aggregated and re-published until it appears to have many independent sources. In reality, every source traces back to the same original statement, which may itself have been speculation, marketing, or error.

AI retrieval systems are vulnerable to this because they rank by frequency and authority signals. If enough high-traffic pages repeat a claim, the system treats it as well-established fact.

Citation Integrity™ treats repetition as a signal to investigate, not a substitute for original evidence.

Knowledge has a half-life. Prices change, regulations are updated, research is superseded, and companies release new products. An answer that was correct six months ago can become quietly wrong today.

The danger is greatest when the answer sounds timeless. A model may present old guidance as current, or quote a study that has since been retracted or contradicted, without any indication that the evidence has aged.

Integrity assessment includes freshness: is this the kind of claim that depends on recent information, and if so, is the source recent enough to support it?

Large language models are trained to be helpful and coherent. That coherence can become overconfidence: a definitive answer where the evidence is contested, incomplete, or genuinely uncertain.

The problem is not that the model has an opinion. It is that the answer does not reflect the distribution of evidence. Dissenting sources, missing data, and low-confidence findings are flattened into a single, polished statement.

Good integrity practice means surfacing uncertainty, not suppressing it. Users and downstream systems need to know when the evidence is mixed.

Not every integrity problem is about bad sources. Sometimes the best source is invisible to automated systems: it lives behind a login, inside a PDF with no structure, in a video without a transcript, or on a page blocked by robots.txt.

When authoritative information is hard to retrieve, models fall back on whatever is available. The result is not a lie, but a substitution: a weaker source replaces a stronger one because the stronger one could not be accessed.

Information owners and AI builders share an interest in making valuable evidence technically accessible without compromising rights or quality.