AI research with citations: a failure-mode taxonomy, not a "use RAG" answer

The original-content test for this topic

Most pages about AI research with citations describe the same problem (“AI hallucinates sources”) and the same solution (“use RAG”). That description is true and incomplete. It treats AI citation reliability as a single binary — hallucinated or not — when the reality has at least five distinct failure modes, each requiring its own verification step.

The honest framing is different: a “cited” AI report can be wrong in five different ways, and a workflow that only checks for one of them is unsafe by default. A page that does not enumerate these failure modes — and tell the reader what to verify against each — is teaching cargo-cult citation discipline. This page treats citation reliability as a taxonomy of failures, not a single problem with a single fix.

The reference data this page anchors to: a 2024 Nature study found LLMs fabricate roughly 36% of generated references; INRA’s published research reports a hallucination spectrum of 17–55% across general-purpose AI tools, dropping below 1% only with multi-layer source validation; and a 2026 Nature analysis warned that “hallucinated citations are polluting the scientific literature” — meaning the failure is no longer hypothetical. These numbers shape what a defensible workflow has to do.

The five failure modes of an AI citation

Treating “hallucination” as a single failure obscures the four other ways an AI citation can be wrong. Each is observable, each has a different cause, and each requires a different check.

A workflow that only catches fabrication (the most discussed failure) still ships briefs with misattribution, misquote, drift, and detachment errors. The Nature 2024 study reported 36% fabrication; misattribution and misquote rates in published audits are lower individually but additive — the combined unreliability of an unverified AI citation is higher than the fabrication rate alone.

Prevention by design vs post-hoc detection

Two architectural approaches dominate the AI-citation space, and they are not equivalent.

Post-hoc detection is the dominant pattern in consumer AI tools. The model generates citations from training data; a separate validator (often the same model or another LLM) checks whether the citations resolve. GPTZero’s hallucination detector, citely.ai, and INRA’s verification layers all sit downstream of generation. They catch fabrication well, misattribution moderately, misquote poorly, drift not at all, detachment never.

Prevention by design is the architectural alternative. The model is not allowed to generate a citation that is not grounded in a fetched, current source at generation time. RAG (retrieval-augmented generation) is the most common implementation: retrieve before generating, attach the citation to the retrieved chunk, and refuse to invent. Perplexity, Bing Chat, Bard, and the more disciplined research-focused tools all use prevention-by-design as the foundation, with detection as a second layer.

The trade-off is real. Detection-only systems are cheaper to build and work with any underlying model. Prevention systems require retrieval infrastructure, source-attribution plumbing, and a refusal mechanism when retrieval fails. For low-stakes use, detection is enough. For research that will be defended — academic, financial, legal, journalistic — prevention is the only architecture that survives audit. A user who does not know which architecture their tool uses cannot calibrate how much to verify.

Three properties make a citation defensible

A citation that survives review has three observable properties. Each is testable; each is missing in chat-style AI output.

Provenance: the citation includes enough metadata to re-find the source independently of the workspace. Not “the model said X” but “Nature, vol 626, Jan 2024, DOI 10.1038/d41586-026-00969-z, retrieved 2026-04-29.” A workspace that hides the metadata behind a hyperlink and loses it on export is failing this test.

Locality: the citation attaches to the specific claim it supports, not to the paragraph or the section. A paragraph with five claims and one bibliography link is unverifiable — a reader cannot tell which of the five claims the link supposedly supports. Paragraph-level citation is a floor, not a ceiling. Sentence-level or claim-level is what audit demands.

Persistence: the citation survives revision. When the paragraph moves between sections, the citation moves with it. When the paragraph splits into two paragraphs, the citation duplicates onto both. When the paragraph is rewritten, the citation either survives the rewrite or surfaces as “needs re-verification.” Workspaces that lose citations on edit produce drafts that look cited and audit as uncited.

The University of North Carolina at Charlotte’s library guide on hallucinated citations and the academic literature on AI citation integrity converge on these three properties as the operational definition of “defensible.”

Citations through editing: the silent failure mode

The most common citation failure in production research is not fabrication. It is detachment during editing. A paragraph is generated with a correct citation. Two weeks later, the user moves the paragraph to a different section, splits it, edits the wording, or merges it with another paragraph. By the time the deliverable ships, the citation graph no longer reflects what the document actually says.

This failure is invisible at the moment it happens. The user sees a hyperlink that still resolves; they assume the citation is intact. The auditor opens the link three months later and finds it supports a different claim than the one in the document.

A useful test: take any paragraph in the deliverable. Does its citation point to a source that contains the specific claim in the paragraph as it currently reads? Not as it was written. As it currently reads. Most AI-citation workflows fail this test on roughly 10–20% of paragraphs after a single round of revision, and the rate grows with each subsequent edit.

The fix is a citation system that treats the citation as part of the paragraph (moves with it, splits with it, surfaces conflicts when the paragraph rewrites past what the source says), not a hyperlink the user pastes into the prose.

What you must verify before reuse

A defensible reuse pass has four steps. Each step catches a different failure mode.

1. Resolve every URL / DOI. This catches fabrication. If the link does not resolve, the citation is invalid regardless of what the surrounding text says.
2. Open the source and confirm the claim is in it. This catches misattribution. The source exists but supports a different sentence — or worse, the opposite of the sentence in the deliverable.
3. For quoted strings, verify verbatim. This catches misquote. AI summaries put paraphrases inside quotation marks more often than is comfortable.
4. Check timestamps for time-sensitive claims. This catches drift. Pricing, customer counts, regulatory status, and current events are valid only as of the retrieval date.

A citation that survives all four steps is defensible. A citation that survives only the first is fabrication-checked but otherwise unverified. The mistake most users make is to assume that a working URL means a working citation.

A note from building Innogath

The five-failure-mode taxonomy did not come from a paper. It came from a manual audit we ran across early Innogath outputs: every citation in real user reports, opened and checked against the underlying source. The most surprising finding was that detachment-during-editing was more common than fabrication — the failure mode SEO blogs do not write about turned out to be the largest in practice. That finding pushed us to build claim-citation binding into the editor, not as a post-hoc check.

Where Innogath fits

Innogath uses a prevention-by-design architecture: retrieval happens before generation, citations attach at the claim level (not paragraph level), and the citation graph survives revision because citations move with their claims through edits, splits, and merges. Drift is handled by per-source freshness windows and re-fetch on use. Fabrication is structurally prevented by refusing to generate a citation when retrieval fails.

For the methodology that uses these citations defensibly, see systematic literature review with AI and AI competitive intelligence. For the broader workflow this sits inside, see the deep research guide.

References

Nature, “Hallucinated citations are polluting the scientific literature” (2026), DOI 10.1038/d41586-026-00969-z. Nature, “Can researchers stop AI making up citations?” (2025), DOI 10.1038/d41586-025-02853-8. INRA.AI’s published research on citation accuracy and the 6-layer validation approach. The University of North Carolina at Charlotte library guide on AI hallucinated citations. The 2024 Nature study cited in INRA’s reporting found a 36% fabrication rate across general-purpose LLM-generated references.

For technical background on retrieval-augmented generation and prevention-by-design architectures, see Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (Facebook AI Research, 2020), and the public technical documentation for Perplexity, Bing Chat, and Anthropic’s Claude on grounded generation.

For adjacent methodology, see branching research pages, systematic literature review with AI, and AI competitive intelligence.