Dozens of Academic Papers Have References That Lead Nowhere

References are supposed to let readers retrace the evidence. Two recent studies found that some of those trails end at papers that cannot be found.

A University of Auckland team reviewed 3,201 peer-reviewed papers published between 2023 and 2025. Automated checks and manual review left 69 papers, or 2.16%, with at least one reference the researchers could not verify. They did not claim that every unmatched reference came from AI.

A May 2026 preprint by researchers from Cornell University, UCLA, Tsinghua University and UC Berkeley looked much further: 111 million references across 2.5 million papers and preprints in arXiv, bioRxiv, SSRN and PubMed Central. The authors compared recent unmatched-reference rates with the years before widespread LLM use. From that increase, they estimated 146,932 hallucinated citations across the four collections in 2025.

What the researchers found

1. Most affected papers looked normal

The large study did not find a small pile of papers in which every reference was false. It found many papers with only a few unmatched references. One plausible-looking citation among dozens of real ones is easy to miss.

2. Peer review did not catch everything

The Auckland study examined peer-reviewed conference papers. Some references still led nowhere. One unverifiable title even appeared in Google Scholar after other publications cited it. Repetition had given the reference a record of its own.

3. A real paper can still be used badly

Both studies mainly asked whether a cited work exists. A real paper can still be misquoted or attached to a claim it never made. The large study identifies this as the harder problem, and one that reference matching alone cannot solve.

What AI FactScan can do

AI FactScan checks citations while they are still inside an AI answer. It grades source types and points out links worth opening before they are carried into an essay, report or manuscript.

The grades turn a long source list into a reading order. An academic database, an official institution and a news article do different jobs. Search results and an AI system citing itself need more caution.

AI FactScan turns the source list into a practical first-pass audit, showing you where to start checking and which references deserve attention first.

A citation can be real, correctly formatted, and still fail to support the claim beside it.

Read the source, then decide

Start with the citation carrying the most important claim. Open the original source, not a search snippet or another summary. Check the title, authors, date and identifier. Then find the passage, table or result behind the answer.

Once the citation has been verified, ask three questions:

1. Does it support the specific claim, or does it merely discuss the same topic?
2. Who and what did the researchers actually study, and where do the findings apply?
3. Does the AI answer sound more certain than the authors' conclusion?

Verification tools can shorten the route. They cannot take the final reading decision away from you.