What Are Hallucinated Citations and Why They End Up in Theses
Hallucinated citations look real. The names sound right, the journal is plausible, the DOI has the correct format. The paper does not exist. Here is what happens and why examiners find them.
The citation looks right. The paper does not exist.
Hallucinated citations are references generated by a language model that contain plausible-looking metadata for a paper that was never written. What are hallucinated citations at the surface? Author names that sound like real people in the field, a journal that publishes work of that kind, a year within the last decade, and a DOI that follows the correct format. What are hallucinated citations underneath? A statistically plausible arrangement of tokens that describes nothing.
The problem is not that the citations are obviously wrong. The problem is that they are subtly wrong in a way that survives skim-reading. A candidate copying references from an AI-assisted draft into a thesis bibliography can pass their own eye check three times and still miss the fabrication. Examiners catch it because they check the specific paper against Crossref, and the paper is not there.
Which two properties collide in a hallucinated citation
A hallucinated citation is what you get when two properties meet in one reference. The first is a convincing form: it is built from real-looking parts in the right citation style. The second is a missing record: no matching work exists.
Either property alone is harmless or easy to catch. A badly formed reference with no record looks wrong at a glance. A convincing reference with a record is simply a citation. The trouble starts only when both are true at once, because the form hides the absence.
The audits show how convincing the form can be. In Walters and Wilder’s data, most fabricated article, book and website citations named real journals, publishers or organisations. Among fabricated journal articles, only 2 percent named a journal that does not exist.
How hallucinated citations differ from other reference problems
Three different problems get called “fake citations”, and each needs a different check. Sorting them first saves you from deleting a real source or keeping an invented one.
| Problem | What is wrong | Typical sign | Check and fix |
|---|---|---|---|
| Hallucinated citation | The work does not exist | Real journal, plausible authors, a DOI that fails or does not match | Search by DOI, title and author; if nothing exists, delete it |
| Miscited source | The work exists, details are wrong | Right title and authors, wrong year or pages | Compare each field with the publisher’s record and correct it |
| Misused source | The work exists but does not say what you claim | A perfect reference on the wrong sentence | Read the passage, then rewrite the claim or change the source |
Only the first row is a hallucination. A reference that passes every registry check can still sit in the third row.
What a hallucinated citation is built from
A hallucinated citation is assembled field by field, and any field can be borrowed from reality. Here is a fill-in template showing where each part can come from. It is an illustration, not a real reference.
(Author A, Author B) (Year). (Title on the right topic). (Real journal name), (Volume)((Issue)), (Pages). https://doi.org/(string in the right format)
- Authors: names that sound right for the field, sometimes real researchers in that area.
- Title: a phrase that matches the topic of your sentence closely, which is what makes it feel relevant.
- Journal: a real journal in most cases, as the 2 percent figure above shows.
- Year, volume, issue and pages: numbers in a plausible range for that journal.
- DOI: a string in the right shape. The shape proves nothing, so open it and compare what loads with every field of the reference.
Because every field can look right, reading a reference closely does not tell you whether it exists. Only finding the record does. Our guide on how to spot hallucinated citations in your thesis turns that into a checking routine for a whole draft.
How often this happens in AI-assisted drafts
Bhattacharyya and colleagues (2023) audited 115 references generated by ChatGPT for medical research topics. Only 7% were fully authentic and accurate. Forty-seven percent were completely fabricated, and the remainder were partially wrong: real authors stitched to nonexistent papers, or real papers cited with wrong details.
Walters and Wilder (2023) ran the same test at scale. Across 636 citations in 84 literature reviews, 55% of GPT-3.5 references were fabricated. GPT-4 fabricated 18% under the same protocol. Newer models do better; none do well enough for a thesis-scale bibliography to be trusted without checking. Our write-up of the Walters and Wilder study sets out its method and every figure.
Chelli and colleagues (2024) tested systematic-review generation across models and found hallucination rates ranging from 28.6% for the best model to 91.4% for the worst. DOI accuracy across the whole sample was 16 to 20%. In short, most DOIs generated by AI tools do not resolve to the paper described.
The most current field-scale data comes from an audit by Topaz and colleagues (2026) in The Lancet. They examined 2.5 million PubMed-indexed papers. One in every 277 papers published in early 2026 contained at least one fabricated reference, up from one in 2,828 in 2023. That is a roughly tenfold rise in three years, tracking the adoption of AI-assisted drafting.
Why language models produce citations that do not exist
A language model has no database of real papers. It has learned what citations look like, so when a sentence calls for a reference it can write one with the right shape and nothing behind it. The full mechanism, including why invented references cluster around niche claims, is in our guide on why AI makes up citations.
How hallucinated citations end up in theses
Most fabricated citations enter a thesis through one of three routes. The first is drafting: a candidate asks an AI tool to suggest references for a paragraph, copies the references into the bibliography, and forgets to verify them. The second is summarising: a candidate asks the tool to summarise the literature on a topic, and the summary includes citations that were never in any paper the model read. The third is translating: a candidate writes a paragraph in one language and asks the tool to render it in another, and the tool adds citations along the way that were not in the original.
None of these routes is malicious. All three produce fabricated citations at scale, because the underlying model does the same thing in each case: it generates plausible-looking references when the context calls for a reference. The candidate then treats the output as reliable, because the output reads as reliable. Which of these uses your university permits at all is a separate question, covered in our guide on whether you can use ChatGPT for your thesis.
This is where a verification step upstream of the bibliography matters. Every citation in a thesis, whether typed by hand or pulled from an AI-assisted draft, has to resolve against a real database entry. It is worth a dedicated pass to check your whole reference list before the defense. Thesisroom’s citation guard looks each entry up in the academic record, drawing on ten registries. Fabricated citations fail the DOI check first, because the DOI does not resolve. Real-but-miscited citations fail the metadata check, because the returned author list does not match. The two failure modes have different fixes, but both surface in the same run.
What examiners do with a fabricated citation
Every viva has a citations question. Sometimes it is direct: “Why did you cite Author X for this claim?” Sometimes it is structural: “Take us to page 47.” A fabricated citation surfaces at exactly the moment the examiner tries to check the source. Park (2003) notes that examiners are trained to spot-check unfamiliar references, especially those doing heavy work in the argument.
The immediate consequence of a fabricated citation is a difficult moment in the room. The larger consequence is the impression it forms. Mullins and Kiley (2002) describe examiners as forming a strong first impression of the thesis in the first hour of reading, and the reference list is one of the three things that shapes that impression. A bibliography with even one fabrication tells the examiner that the rest cannot be trusted without checking. That check happens whether the candidate notices or not. If a query about AI use does follow, it helps to know how to prove your thesis is your own work from the paper trail you already keep.
What Thesisroom does with this
Thesisroom’s citation guard extracts every bibliography entry and looks it up in the academic record, drawing on ten registries. Every entry returns as verified, flagged, or unverifiable, and every flag names the specific database check that produced it. The full method treats verification as a whole-degree habit: the reference list that leaves a candidate’s literature review clean is the same one that will not surface a fabrication in the room.
The tool does not audit the fourth step, whether a real source supports a specific claim. That is a reading task the writer has to do. It does flag every entry that could not be verified in any database, so the reading is targeted at the sources most likely to have been miscited.
Frequently asked questions
Are hallucinated citations always from AI tools?
Not always, but the vast majority now are. Before language models were widely used, fabricated citations existed but were rare and usually deliberate. Since 2022, the base rate has risen sharply, tracked in Topaz and colleagues’ 2026 Lancet audit as a roughly tenfold rise in fabricated references in PubMed papers between 2023 and 2026.
Can I tell a hallucinated citation just by looking at it?
Sometimes, but not reliably. The author names are usually plausible, the journal exists, and the year is recent. The DOI is often the tell: it follows the correct format but does not resolve when you paste it into https://doi.org. Some fabricated DOIs resolve to a different paper entirely, which is even harder to spot.
Do only bad models hallucinate citations?
No. Even the best current models fabricate citations at meaningful rates when asked for references. Walters and Wilder (2023) found GPT-4 fabricated 18% of citations across their sample, down from 55% for GPT-3.5. Rates have fallen over time, but not to a level that makes verification optional.
What percentage of AI-generated citations are fake?
Rates vary widely by model and topic. Bhattacharyya and colleagues (2023) found 47% completely fabricated in ChatGPT-generated medical references. Walters and Wilder (2023) found 18% to 55% depending on model. Chelli and colleagues (2024) found rates from 28.6% up to 91.4%. The base rate is high enough that every AI-assisted bibliography needs verification.
What happens if an examiner finds a fabricated citation?
At minimum, it becomes the topic of extended discussion in the viva. At worst, it triggers a formal academic misconduct process, depending on the institution’s policy and the candidate’s account of how it entered the bibliography (Park, 2003). Owning the origin (“I used an AI tool during drafting and did not verify”) is usually better than pretending the citation was real.
How do I catch hallucinated citations across a whole bibliography?
Manually, you resolve every DOI and cross-check the metadata against the citation. At scale, tools that run the same checks against Crossref, OpenAlex, and Semantic Scholar are much faster. Citation Check runs the full sweep against the registries and flags anything that fails, with the record behind each verdict.
One takeaway before you close this tab
The most useful mental model for a hallucinated citation is that it is what a language model produces when it does not know the answer. It is not lying. It is completing a pattern. The candidate’s job is to close that gap by verifying every citation before it enters the thesis. When you are ready to run the check at bibliography scale, Thesisroom’s citation guard does the same lookups you would do by hand, across the whole bibliography at once.
References
Bhattacharyya, M., Miller, V. M., Bhattacharyya, D., & Miller, L. E. (2023). High rates of fabricated and inaccurate references in ChatGPT-generated medical content. Cureus, 15(5).
Chelli, M., Descamps, J., Lavoué, V., Trojani, C., Azar, M., et al. (2024). Hallucination rates and reference accuracy of ChatGPT and Bard for systematic reviews. Journal of Medical Internet Research, 26.
Mullins, G., & Kiley, M. (2002). ‘It’s a PhD, not a Nobel Prize’: How experienced examiners assess research theses. Studies in Higher Education, 27(4), 369-386.
Park, C. (2003). Levelling the Playing Field: Towards Best Practice in the Doctoral Viva. Higher Education Review, 36(1), 47-67.
Topaz, M., et al. (2026). Analysis of fabricated references in PubMed-indexed literature. The Lancet.
Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13.