The reference list runs to 42 entries. A reviewer picks five at random to follow up and cannot find two of them. The first author is real, a well-known name in the field. The journal is real. The titles sound exactly on topic. The only problem is that those papers were never written. The manuscript's author was not trying to deceive anyone: he had asked an AI assistant to “suggest a few more sources for this paragraph” and pasted the result in.
For an editor or examiner, the reason changes little. A non-existent source in your paper signals that you have not read what you cite.
Why AI assistants invent citations
Language models generate text by predicting likely sequences of words. Academic citations follow a very regular pattern: names, year, title, journal, volume, pages. Models learn that pattern extremely well, and when asked for a source supporting a claim, they produce something that looks like a correct citation, assembled from fragments that often appear together. Unless the tool is built to search and return real records, no database is consulted.
Even tools with search built in make a different error: citing a real paper while attributing to it something it does not say.
Three kinds of AI citation error
| Type | Description | How hard to catch |
|---|---|---|
| Fully fabricated | No such paper exists | Easy, if you look |
| Chimera | Title from one paper, authors from another, year from a third; the DOI resolves to something unrelated | Moderate |
| Real but misrepresented | The paper exists but does not support the claim attached to it | Hard; requires reading |
The third type is the most dangerous because it passes every automated check. Only reading reveals it.
Quick warning signs
- The title matches your sentence a little too perfectly, as if written for that paragraph.
- The author is prominent in the field, but the topic sits outside their known research line.
- Volume and year do not fit the journal's publication history (volume 45 of a journal founded a few years ago).
- The DOI looks valid but resolves to a different paper or an error page.
- None of the same author's later papers cite this supposedly important work.
- Page ranges make no sense, such as a two-page empirical study.
A verification routine for each source
- Resolve the DOI at doi.org. If a paper appears, compare title, authors and year with your citation. All three must match.
- No DOI: search the exact title in quotation marks on Google Scholar, Crossref Search or your field's database (PubMed, ERIC, Scopus, Web of Science).
- Check the journal's site: go to the relevant volume and issue and look for the paper in the table of contents.
- Books and chapters: check a library catalogue or the publisher's site, including editors and edition year.
- Read at least the abstract and the relevant passage to confirm the paper says what you claim.
- Import into your reference manager by DOI rather than typing or pasting from AI output. Zotero, Mendeley and EndNote pull metadata from the source, reducing transcription errors.
For sources already in your library, step five is enough. For anything newly suggested by AI, do all six. It averages about five minutes per source, far less than the time spent explaining yourself if a fake reference is found.
Using AI to find sources more safely
- Ask the AI for keywords, concepts and method names to search, then search the databases yourself. Do not ask “give me five references on…”.
- If you use an AI-powered academic search tool, open each link and read the paper rather than trusting the tool's summary sentence.
- Start from a trustworthy review you have read and follow its references and the papers that cite it.
- Never let an AI reformat your reference list from memory; export it from your reference manager.
If you are the marker or reviewer
Spot-checking five to ten references, prioritising those that support the main argument, is an efficient approach. If a source cannot be found, describe the facts (“this item could not be located by DOI or title”) and let your institution's or journal's process take over. A fabricated reference is not by itself proof of AI use; people also copy errors from secondary sources.
If you have already submitted work with fake sources
- Before grading or review is complete: contact whoever is responsible, explain that you found inaccurate references and ask to submit a corrected version. Self-reporting always beats being caught.
- After publication: contact the journal to correct the record. If the fake source was peripheral, a correction is usually enough; if it supported a main conclusion, the argument needs reassessing.
Preventive habits
- Cite only items that are in your reference manager and that you have opened and read.
- Tag each item as “read in full” or “abstract only”.
- Before submission, filter for recently added items and check them one last time.
Your next step: open the reference list of your latest draft, pick the ten entries you remember least clearly and run them through this routine. If all ten pass, your habits are sound; if not, you have just avoided an awkward conversation with an editor.
Câu hỏi thường gặp
Why does AI make up fake references?
Language models generate text that has the form of a citation by combining fragments that often appear together, rather than looking anything up. The result looks plausible but may not exist.
How do I check if a citation is real?
Resolve the DOI at doi.org and compare title, authors and year. Without a DOI, search the exact title on Google Scholar, Crossref or a subject database, then read the relevant passage to confirm the content.
Is citing a hallucinated reference academic misconduct?
It is usually treated as an integrity breach because it shows you cited something you did not read and put false information into the record. The response depends on local rules and whether you reported it yourself.
Are AI search tools safe for finding references?
Safer, because they link to real papers, but they can still misstate what a source says. Open each paper and read the relevant section before citing it.
What should I do if I find fake references after submitting my thesis?
Contact your supervisor or the responsible office straight away, explain the problem and ask to submit a corrected version. Early self-reporting is viewed far more favourably than discovery by others.
