01 · The Question
Can AI Create a Fake Citation Using Details From Real Research Papers?
You ask generative AI for an academic reference. The title sounds familiar, the authors are established researchers, and the journal is legitimate. Even the DOI works. Yet when you open the DOI, you discover that it belongs to a different article.
A closer investigation reveals something unusual. The title resembles one publication, the authors belong to another, and the journal details appear to have come from a third source. Each component looks credible on its own, but the complete reference does not identify a single published work.
How can generative AI produce such a citation, and why might this type of error be harder to detect than a completely invented article?
03 · What You Need to Know
How AI Can Assemble a Convincing but Nonexistent Academic Reference
What Is a Composite or Hybrid Academic Reference?
A composite reference is a bibliographic entry that combines information belonging to different sources into a single, inaccurate citation. The terms composite citation and hybrid citation are useful descriptions of this error pattern, although they are not universally standardized categories in scholarly publishing.
For example, a reference might contain the title of one genuine paper, the authors of another, and the DOI of a third. Alternatively, an AI system might generate a new title that resembles several existing publications while attaching authentic bibliographic details to it.
The central problem is not necessarily that every component is fabricated. It is that the components do not identify the same scholarly object.
Completely fabricated reference
The cited publication cannot be established, and its bibliographic details may be invented.
Composite or hybrid reference
Some details may correspond to genuine publications, but their combination does not accurately identify one work.
Inaccurate reference to a real paper
The intended publication is identifiable, but one or more bibliographic fields are incorrect.
These categories can overlap. A composite citation may also be classified as fabricated if no publication corresponds to the complete reference. Conversely, an incorrect author or DOI may be a simple metadata error rather than evidence that several source records were combined.
Why Can Language Models Mix Bibliographic Information?
Generative language models learn statistical relationships among words, names, titles, and other information. They can reproduce patterns associated with academic referencing without necessarily retrieving a complete bibliographic record from an authoritative database.
When asked to generate a citation, a model may associate a research topic with particular scholars, journals, and publication terminology. If those associations are not grounded in a verified source record, the generated details may be individually plausible but collectively incorrect.
For instance, two researchers may publish frequently on educational technology. A language model might associate one researcher's article title with the other's name because both are strongly associated with the same topic.
This is one possible explanation for a composite-looking reference, not proof of the internal process that produced a particular citation. Without examining a system's generation or retrieval records, researchers generally cannot determine exactly how an erroneous combination arose.
The broader explanation of why generative AI invents academic references helps explain why plausible bibliographic language should not be mistaken for verified publication metadata.
What Does Research Establish About Mixed or Incorrect AI Citations?
Walters and Wilder (2023), in a study published in Scientific Reports, evaluated 636 bibliographic citations generated by GPT-3.5 and GPT-4 across 42 multidisciplinary topics.
They reported that 55% of citations generated by GPT-3.5 and 18% generated by GPT-4 were fabricated. Among citations corresponding to genuine publications, substantive bibliographic errors appeared in 43% of GPT-3.5 references and 24% of GPT-4 references.
The investigation demonstrates that AI-generated citations can contain fabricated publications and incorrect bibliographic details. However, these overall figures should not be presented as the measured prevalence of composite references specifically. The study did not establish that every inaccurate citation resulted from combining multiple real papers.
Research on citation fabrication therefore supports the need to check the relationships among bibliographic fields, while the precise frequency and mechanisms of composite citations require more targeted investigation.
Which Parts of a Citation Can Be Combined Incorrectly?
Almost any identifying field can be associated with the wrong publication. Some mismatches are more consequential than others because they may cause the reader to retrieve an entirely different source.
| Bibliographic Field |
Possible Mismatch |
Verification Method |
| Article title |
A title from one paper is paired with another paper's authors. |
Search the exact title and inspect the publisher's author list. |
| Author names |
Researchers associated with a related topic are attributed to the wrong article. |
Compare the complete author list and sequence with the publication record. |
| Journal title |
A real article is assigned to a different legitimate journal. |
Check the publication venue identified by the publisher. |
| Publication year |
The date of another work or publication event is attached to the citation. |
Examine the version of record and relevant publication dates. |
| Volume, issue, or article number |
Numerical details belong to another publication. |
Compare the complete citation with the publisher's metadata. |
| DOI |
A registered DOI identifies a different article. |
Resolve the DOI and compare the destination record. |
A reference does not become reliable merely because its journal exists or its authors are recognizable. All relevant details must be consistent with the same identifiable publication.
Why Is a Real DOI Particularly Misleading in a Composite Reference?
A DOI provides a convenient route to a registered scholarly object. However, a valid DOI can be attached to an incorrect citation.
Suppose an AI-generated reference describes a paper about generative AI and student learning. Its DOI resolves successfully, but the destination identifies a different study about educational assessment.
The DOI is real. The cited article may also be real under another identifier. Nevertheless, the combination is inaccurate.
This is why finding a real DOI does not validate the surrounding citation. DOI resolution establishes what the identifier points to, not whether the AI paired it with the correct metadata.
Crossref and other DOI registration agencies provide metadata that can assist with these checks. Their records should be compared with the publication itself when discrepancies arise.
How Can You Detect a Composite Citation?
The most useful approach is to treat a citation as a connected record rather than a collection of independent facts.
Begin with the most distinctive identifier, usually the DOI or exact article title. Locate the publication and record its authoritative metadata. Then compare the remaining fields against that same record.
Identify the intended publication.
Search the exact title or resolve the DOI. Determine which scholarly object the available evidence actually identifies.
Compare the complete metadata.
Check the authors, journal, year, volume, issue, and other identifying fields against the same publication record.
Investigate mismatched elements.
Search an incorrect author-title combination or DOI separately to determine whether the disputed details belong to another publication.
Classify the error cautiously.
If different fields correspond to different genuine works, the citation may be composite. If the origin of a field cannot be established, describe it as incorrect rather than assuming where it came from.
Reconstruct only from verified records.
Use the authoritative metadata for the intended publication. Do not assemble a corrected reference from guesses or unrelated search results.
The distinction between identifying an error and explaining its origin is important. A researcher may establish that a DOI belongs to the wrong article without being able to prove that the AI copied it from that article during generation.
Can Retrieval-Enabled AI Systems Produce Composite References?
Yes, although retrieval from structured scholarly records can reduce opportunities for some kinds of fabrication.
A retrieval-enabled system may identify genuine publications but still associate a title with the wrong record, merge information from multiple search results, or generate a citation that does not faithfully reproduce the retrieved metadata.
Similarly, a model processing several uploaded articles may confuse their authors, findings, or publication details.
Structured metadata exports from scholarly databases and reference managers are generally preferable to asking a model to reconstruct bibliographic information from memory. Even then, the final citation should be checked.
Why Does This Matter Beyond Reference-List Accuracy?
A composite reference can interfere with source retrieval, misattribute scholarly credit, and create an inaccurate connection between a manuscript's claim and the research literature.
For example, a reader may follow the DOI and find a paper that does not correspond to the cited title. Alternatively, searching the title may lead to a genuine paper written by different researchers.
The problem becomes more serious if the manuscript also attributes findings to the composite reference. In that case, it may be unclear which publication, if any, actually supports the statement.
Researchers should therefore distinguish bibliographic correction from evidentiary verification. Even after reconstructing a valid reference, they must examine whether the intended source supports the claim.
04 · A Practical Example
How Three Genuine Bibliographic Details Can Produce One False Citation
Hypothetical Example
An AI-Suggested Paper About Generative AI in Higher Education
Suppose an AI assistant provides a reference titled Generative AI Feedback and Undergraduate Writing Achievement, attributed to two researchers who have published on educational technology. It identifies a legitimate education journal and supplies a DOI.
You investigate the citation and discover that the DOI belongs to a different paper about digital assessment. The named researchers have published related work, but neither authored an article with the stated title.
All titles, author associations, and bibliographic relationships in this scenario are illustrative. The example does not describe an actual fabricated reference.
Check the DOI.
The identifier resolves, but the destination publication has a different title and author list.
Search the claimed title.
No independently verifiable publication matches the exact title and named authors.
Investigate the author names.
The researchers are genuine scholars with publications on related subjects, but their records do not establish the claimed article.
Compare the journal information.
The journal exists, but its records do not establish the particular publication described by the AI.
Decide whether to cite.
The combined reference cannot be verified as a single scholarly work and should not be included in the manuscript.
Notice that the individual checks reveal several genuine elements. Yet none establishes the complete citation.
The decisive question is whether the author-title-journal-identifier combination corresponds to one publication. In this scenario, it does not.
06 · What This Means for You
Verify the Relationship Between Citation Fields, Not Just Their Plausibility
When checking AI-generated references, the important unit of verification is the complete publication record.
Instead of asking whether an author exists, whether a journal is legitimate, and whether a DOI works, determine whether those details belong together.
A simple decision framework
If the DOI and title identify different publications
Do not accept the citation. Determine which publication was intended and obtain its authoritative metadata.
If the title exists but the authors are incorrect
Correct the authors using the publisher's record rather than selecting names from related publications.
If multiple fields appear to come from different sources
Treat the entry as potentially composite and investigate each mismatch.
If no complete publication record can be established
Exclude the reference until its identity can be independently verified.
If a valid publication is identified
Use its verified metadata and separately examine whether it supports the intended claim.
For larger bibliographies, importing records directly from publishers, scholarly databases, or reference managers can reduce the need to reconstruct citations manually.
The broader verification of AI-generated academic citations should include checking that all bibliographic fields identify the same work.
