03 · What You Need to Know
How Can AI Create a Research Paper That Never Existed?
A Fabricated Reference Does Not Need Fabricated Ingredients
The simplest hallucinated citation would contain a fictional author, fictional article, and fictional journal. Such a reference may be relatively easy to expose.
Generative AI can produce something much more convincing.
It can name a real researcher whose work concerns your topic. It can place the invented article in a real journal that publishes that kind of research. It can generate a title using terminology common in the field. It can then add realistic publication metadata.
The resulting citation may be false even though several individual components are true.
Real components
The author exists. The journal exists. The research topic is appropriate.
Real publication
Those authors actually published that specific work under that title in that venue.
The first does not establish the second.
Research Has Documented This Mixing of Real and Fabricated Details
Walters and Wilder examined 636 citations generated by GPT-3.5 and GPT-4 for short literature reviews. Under their experimental conditions, 55% of GPT-3.5 citations and 18% of GPT-4 citations referred to fabricated works.
Importantly, many of the fabricated article, book, and website citations included names of real journals, publishers, or organizations. The fabricated reference therefore did not necessarily announce itself through obviously fictional entities.
A separate study by Chelli and colleagues examining references generated for systematic-review tasks similarly documented hallucinated papers with titles and journal information resembling genuine literature. Their analysis included examples in which hallucinated references bore similarities to authentic papers in titles and author lists.
These studies concern particular model versions and experimental tasks rather than all current generative AI systems. They nevertheless demonstrate the underlying failure mode: bibliographic plausibility can be constructed from genuine scholarly patterns and entities.
AI Can Invent an Article Title That Sounds Almost Inevitable
Academic article titles are patterned. They frequently combine a topic, population, method, relationship, intervention, outcome, or setting in recognizable ways.
If you ask for research on a very specific combination of concepts, a language model can generate a title that sounds exactly like the paper you hoped someone had written.
For example, a title might combine the name of an established theory, a familiar educational technology, a university population, and a commonly studied outcome. Every phrase may be academically appropriate. The complete title may nevertheless identify no publication at all.
This creates a particular danger during literature searching: the generated title may be so semantically well matched to your question that you assume it was retrieved rather than generated.
Watch Out
A title that perfectly matches your research question deserves verification, not celebration. Generative AI is very good at producing the title that a relevant paper would have if such a paper existed.
AI Can Use the Names of Real Researchers Incorrectly
Author names can create a strong credibility signal, especially when you recognize them.
Suppose an AI attributes a paper about a theory to a prominent scholar associated with that theory. The attribution feels intuitively right. That researcher may indeed have written extensively on the subject.
But expertise in the topic does not establish authorship of the specific paper.
A model can attribute a nonexistent publication to a real scholar, combine researchers who never coauthored a paper, omit genuine coauthors, or add someone who was not involved.
For common surnames and initials, another problem arises: researchers with similar names can be confused with one another. Author verification may therefore require checking institutional affiliations, ORCID records, publisher profiles, or the authors' actual publication records rather than merely finding someone with the same name.
AI Can Invent a Journal Name
Journal titles also have recognizable linguistic conventions. Terms such as International Journal of, Journal of, Research in, Advances in, and Review of can be combined with disciplinary terminology to create names that sound entirely plausible.
A generated journal may therefore resemble an established publication while not existing at all.
Alternatively, the model may slightly alter the title of a genuine journal. A single added or substituted word can turn a legitimate journal name into a nonexistent one while preserving the appearance of authenticity.
Verify the exact journal title through the publisher, library catalog, relevant indexing database, ISSN Portal where appropriate, or another authoritative source.
A Real Journal Does Not Verify the Article
Suppose you check the journal and discover that it exists. That is useful, but you have verified only the venue.
Walters and Wilder's findings illustrate precisely why this matters: fabricated references can contain real journal names.
You still need to establish that the journal actually published the article in question.
Journal check
Does the publication venue exist?
Article check
Does the exact article appear in that journal's publication record?
Authorship check
Are the claimed researchers actually listed as authors?
Metadata check
Do the year, volume, issue, pages or article number, and identifier correspond to the same work?
AI Can Slightly Alter a Real Article Title
Not every title error creates a wholly nonexistent paper. A model may paraphrase, truncate, expand, or otherwise alter the title of a genuine publication.
This can make verification confusing. Your search may return a very similar paper, and you may reasonably wonder whether the AI merely made a small transcription error.
Sometimes that is exactly what happened. But similarity should not be used to silently repair a citation.
Confirm the official title and reconstruct the citation from the verified publication record. If several details differ, consider whether the AI has combined information from multiple real studies into a nonexistent paper.
Real Authors, Real Title Words, and a Real Journal Can Still Produce a Fake Paper
Consider how convincing the following pattern can become:
| Component |
Status |
| First author |
Real researcher in the field |
| Coauthor |
Real researcher in a related field |
| Article title |
Generated from genuine disciplinary terminology |
| Journal |
Real journal publishing on the topic |
| Year |
Plausible |
| DOI |
DOI-shaped but unverified |
| Complete publication |
Nonexistent |
A researcher checking only the first author and journal might conclude that the citation is credible. Verification needs to test whether all the pieces converge on one identifiable publication.
A Real Paper Can Still Have Incorrect Authors or Journal Metadata
The opposite problem also occurs. The paper itself may be genuine while some generated bibliographic elements are wrong.
Walters and Wilder found substantive citation errors among references to real works, including errors involving author names, article titles, dates, journal titles, volume and issue information, page numbers, and publishers.
This is why the broader question of AI-hallucinated citations and references cannot be reduced to a binary “real paper versus fake paper” test. Bibliographic accuracy exists at the level of individual fields as well.
Authoritative Records Are Better Than Plausibility Checks
Trying to decide whether a publication “sounds real” is a weak verification method. Search for it.
Depending on the source, useful verification routes include the publisher's website, Crossref, PubMed, discipline-specific scholarly databases, library catalogs, indexing services, ORCID for author identity, and other authoritative records.
Exact-title searching can be particularly useful. If an AI gives you a complete title, search that title as a phrase and then verify any result against the remaining metadata.
If you find only similar titles rather than the claimed work, investigate rather than assuming the AI was close enough.
