How to Check Whether a Citation Actually Exists

A few years ago, a fake citation in a draft usually meant a typo, a misremembered year, or a reference copied from someone else's bibliography without checking. Today there's a new and much more convincing source: AI chatbots that, when asked for references, will sometimes produce ones that were never written. The author names are real, the journal is real, the title sounds exactly like something that journal would publish. The paper just doesn't exist.

This happens because a language model generates text that fits the pattern of a reference, not text retrieved from a database of real ones. A plausible citation and a true citation look identical on the page, which is precisely what makes the problem easy to miss. It isn't only a student problem either. In 2023, lawyers in a US federal case were sanctioned after filing a brief that cited court decisions ChatGPT had invented, and publishers have since had to correct or retract academic work over references that turned out not to exist.

The good news is that a fabricated reference almost never survives a few minutes of checking. Here's the routine.

The four ways a reference can be wrong

It helps to know what you're looking for, because "fake" covers more than one failure.

The paper doesn't exist at all. The title, authors, and venue are stitched together from things that do exist, but no such paper was ever published.

The paper exists, but the details are wrong. Right title, wrong year. Right authors, wrong journal. These are the most common, and they're also the easiest to overlook because a quick search seems to confirm the reference.

The DOI points somewhere else. A DOI that looks valid resolves to a completely different paper, or to nothing. This is a strong red flag, since DOIs are exactly the detail people assume is machine-checked.

The paper exists and is correct, but it doesn't say what you're citing it for. This one isn't technically a fake citation, but it's the most damaging in peer review, because it means the claim in your text is unsupported. It's also a mistake humans make without any AI involved.

A five-minute check for each reference

1. Search the exact title in quotes. Put the full title in quotation marks in Google Scholar, Semantic Scholar, or OpenAlex. A real published paper will almost always come up immediately. If nothing matches, try just the most distinctive part of the title before concluding anything, since titles occasionally differ slightly between preprint and published versions.

2. Resolve the DOI yourself. Paste the DOI after https://doi.org/ in your browser and check that the page it lands on matches the title, authors, and year in your reference. Don't just check that the link works. Check that it goes to the right paper.

3. Confirm the venue actually published it. Look up the journal's volume and issue, or the conference proceedings, and see if the paper is listed there. Fabricated references often have page ranges or volume numbers that don't correspond to anything.

4. Read the abstract against your sentence. Put your claim next to the paper's abstract and ask whether a skeptical reader would agree that one supports the other. If the paper is about a related question, a different population, or reaches a more cautious conclusion than your sentence does, you either need a different source or a more carefully worded claim.

5. Check whether it's been retracted. A real, correctly cited paper can still be a liability if it's been withdrawn. The Retraction Watch database is free to search, and many publishers mark retracted papers prominently on the article page.

When to be especially suspicious

Some patterns should make you check a reference first rather than last. Be wary of references you can't remember finding yourself, anything that came from a chatbot conversation, references that fit your argument unusually perfectly, and citations with an author you know but a title you've never seen from them. None of these mean the reference is fake. They just mean it's the one to verify before you spend time on the others.

Starting from real papers instead

The underlying fix is to reverse the direction of the search. Instead of starting with a reference and trying to confirm it exists, start from databases that only contain papers that do exist, and find the ones that support your claim.

That's the approach Research Verifier takes. It never generates references. Every result comes directly from Semantic Scholar, OpenAlex, Europe PMC, PubMed, arXiv, or your own uploaded papers, and exports include the DOI wherever one is available, so steps 1 through 3 above are already answered by where the paper came from. Step 4 is still yours: results are ranked by how closely a paper's abstract matches your claim in meaning, but reading that abstract before you cite it is the part no tool should do for you.