A citation makes two claims: this document exists, and it says this. An AI-generated reference can fail either one. The paper may be invented. Or it may be real and say something other than what the report claims. A link that opens settles neither question on its own: it shows that something sits at that address, not that it is the cited work or that it supports the sentence.
Start with the three claims your conclusion depends on. Open each source yourself, confirm it is the cited work, find the passage and read what it supports; then set the most important claim against independent evidence. Asking the tool whether its source is real replaces none of these steps. The check below leaves you with a record of what you checked and what is still open.
A fabricated case, vouched for by the same tool
A passenger sued Avianca in New York, saying a metal serving cart had struck his knee on a flight. On 1 March 2023 his lawyers filed a response in federal court that cited six decisions generated by ChatGPT. None of the six existed. One was Varghese v. China Southern Airlines.
The sanctions order records two failures that matter for any citation check. First, the fake Varghese carried a Federal Reporter citation that belonged to a different, real case. A reference can look right, and even point at a real document, without identifying the work it names.
Second, one of the lawyers asked ChatGPT whether the case was real. According to his declaration of 6 June, quoted in the order, ChatGPT answered that Varghese “does indeed exist” and could be found on Westlaw and LexisNexis, two legal research databases. At the sanctions hearing he told the judge: “I just never thought it could be made up.”
On 22 June 2023 Judge P. Kevin Castel imposed a $5,000 penalty on the two lawyers and their firm, and ordered them to send letters to their client and to each judge falsely named as the author of a fake opinion. The order places the duty with the lawyers: “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.” It also notes that the lawyers kept standing by the fake cases after they were questioned; had they come clean then, “the record now would look quite different.”
This is not legal advice. The practical lesson is narrower: the tool that supplied a citation cannot vouch for it, however often it repeats that the source is real.
How often do AI tools make up citations?
There is no single rate. The two studies below measured particular models on particular dates, topics and prompts. Neither gives a current error rate for the tool you use.
Walters and Wilder, 2023. The researchers asked GPT-3.5 and GPT-4 to write short literature reviews on 42 topics in the first week of April 2023, then checked all 636 citations in the 84 papers. Their Scientific Reports paper separates invented works from errors in references to real ones.
| Walters and Wilder (2023) | GPT-3.5 | GPT-4 |
|---|---|---|
| Works cited | 222 | 414 |
| Fabricated (the work does not exist) | 55% | 18% |
| Real works cited | 101 | 340 |
| Of those real works, citations with substantive errors | 43% | 24% |
The last row counts errors in the reference itself, such as wrong authors, dates, volume or page numbers. It does not measure how often a text misstates what a real paper found. One more figure from the same sample: 70% of the book chapters GPT-4 cited were fabricated.
Linardon and colleagues, 2025. The team asked GPT-4o in June 2025 for six literature reviews on mental-health topics and checked all 176 citations. In their JMIR Mental Health study, 35 were fabricated. Of those, 33 carried a DOI, the persistent identifier used for scholarly papers and other objects. Twenty-one of those DOIs resolved to unrelated articles; the other 12 were invalid.
So a DOI that opens proves only that the identifier points somewhere. You still have to confirm that it is the cited work, and then what that work supports.
Why does an AI invent a source?
A language model can generate a plausible citation, complete with author, year, journal and page numbers, without a document behind it. The familiar format is no evidence that the source exists. Confident false statements like these are commonly called hallucinations.
A 2025 paper by Adam Tauman Kalai and colleagues describes the problem: “Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty.” The authors argue that training and evaluation procedures reward guessing over acknowledging uncertainty.
The same applies when you ask the model to confirm its own citation: the confirmation is more generated text, not a check. Walters and Wilder note that ChatGPT “often provides inaccurate responses when asked to verify the legitimacy of the works it cites.” The Avianca lawyer got exactly that kind of answer. Open the document yourself.
The 15-minute check
Fifteen minutes is a planning time box for a first pass on the citations your conclusion stands on, not a measured completion time. Set a timer. Whatever you cannot finish stays marked as unverified.
Minutes 0 to 2: pick three. Choose three claims that would change the decision if they were false. These are load-bearing claims. Copy each claim as written and note its citation. If more claims carry the decision, put them next in the queue.
Minutes 2 to 11: three minutes per citation, four questions.
- Does the identifier open a document? Paste the DOI after
https://doi.org/, or open the source link. A working address is a lead. A dead link leaves the source unverified; it does not by itself prove fabrication. - Is it the same work? Compare title, authors, year and journal with the citation. If you land on a different paper, the citation fails here, as the 21 unrelated DOIs in the 2025 study would have. If the DOI is missing or broken, search the exact title on the publisher’s site or in Crossref Metadata Search. Record bibliographic errors even if you locate the intended work.
- Does the cited passage exist? Open the full text and search for the key number or phrase. Read the surrounding text and any relevant tables. An exact quote must appear word for word. If you cannot access or locate the passage, leave the check unresolved.
- Does it say what the report says? Compare the population, period, number, direction and qualifications. “May reduce” is not “reduces”. A finding about students does not establish the same result for managers. Record what the source supports, then narrow the report’s claim if necessary.
Minutes 11 to 15: compare independent evidence. For the most important claim, look for a second source and check where its evidence comes from. A different organization may still be using the same data, and two articles repeating one study are one evidence base. Record whether the second source supports, qualifies or contradicts the claim. If you cannot find one, record that gap.
| Check | What to establish | What to record if unresolved or wrong |
|---|---|---|
| Identifier | Opens a source record or document | Broken link; source not yet located |
| Same work | Title, authors, year and venue match | Wrong work or bibliographic errors |
| Passage | Number or quote located in context | Passage absent or inaccessible |
| Claim | Scope and wording support this claim | Claim wider, stronger or different |
| Second source | Independent evidence compared | Shared evidence, disagreement or a gap |
Write one line per claim: the claim, source, passage location, checks completed and unresolved issues. That is the start of an evidence ledger.
Remove or correct unsupported claims. Mark unresolved ones as unverified and hold any conclusion that depends on them. If the report has already gone out, send a short correction stating what changed and whether the conclusion changes. Do not quietly replace the file.
Where does TRACE fit?
The check borrows three steps from TRACE, the authorial framework I describe in The Verified Research Loop. Its five steps:
- T · Target. Write the research brief first: the decision, the bounded question, scope, exclusions, deadline and stop rule, before a single search runs.
- R · Retrieve. Obtain the actual source documents, and log what stayed inaccessible, instead of trusting one confident pass.
- A · Assess. Appraise what each source can bear for this claim.
- C · Cross-check. Record one relation per source for each claim and map shared origins. Corroboration is not a document count.
- E · Export. Hand over a versioned decision memo, traceable to the claim ledger, that a co-founder or investor can act on without redoing the work.
In the check above, questions 1 and 2 are Retrieve work: they establish that you hold the cited document. Questions 3 and 4 are Assess, and the comparison with independent evidence is Cross-check. The book puts the rule this way: “treat a model’s citation list as a queue of leads rather than a bibliography.” TRACE is an authorial method with no established accuracy effect. Completing its steps does not guarantee that you found every error or reached the right conclusion.
Evaluate the reasoning, not the fluency covers the same problem from another angle: a fluent answer is not evidence. What is LLM grounding? explains why a valid source link is not enough on its own.
Try this today (15 minutes)
Open the last AI-assisted report, deck or memo you sent. Choose its three most consequential claims and run the check. Keep the source passage beside each claim so another reader can inspect the connection.
At the end, record what you verified, correct what the sources do not support and list what still needs checking. Three checked citations are a useful record, not clearance for the rest of the report.
Cite this:AI fake citations: a 15-minute check for the claims that matter.Len P. van der Hof. https://lenvanderhof.com/en/blog/ai-fake-citations-check/ ·