AI Hallucinated Citations: The Growing Problem in Academic Writing

The Rise of Fabricated References

Every semester, professors discover the same unsettling pattern: student papers citing journal articles that cannot be found in any database. The DOIs lead nowhere. The journal names exist, but the specific articles do not. The author names belong to real researchers, but they never wrote the papers being cited.

These are AI hallucinated citations — references generated by large language models that mimic the format and style of real scholarly work while pointing to publications that have never existed.

What Are AI Hallucinated Citations?

When a language model like ChatGPT or Claude is asked to provide references for a claim, it does not search academic databases. Instead, it predicts what a plausible citation might look like based on patterns in its training data. The result is a reference that has all the hallmarks of a real publication — a convincing author name, a legitimate-sounding journal title, a properly formatted DOI — but corresponds to no actual paper.

This phenomenon is not a bug that will be patched away. It is a fundamental property of how generative models produce text: they optimize for plausibility, not for factual accuracy.

How Prevalent Is the Problem?

Research published in 2024 and 2025 has begun to quantify the scale:

• A study in Scientific Reports found that up to 36.0% of references generated by ChatGPT were entirely fabricated, with fabrication rates varying significantly by discipline (Walters & Wilder, 2023)
• Separate research published in Cureus confirmed that ChatGPT generates citations that appear authentic but correspond to no real publication, with hallucinated references spanning multiple medical specialties (Alkaissi & McFarlane, 2023)
• An analysis in the Journal of Medical Internet Research demonstrated that AI can produce entirely fraudulent scientific articles with convincing but non-existent references (Májovský et al., 2023)

The problem is amplified by the fact that hallucinated citations are designed to look correct. A human reviewer scanning a reference list will often not catch them without individually verifying each entry against a database.

Notably, newer models with retrieval-augmented generation (RAG) capabilities have reduced but not eliminated the issue. Even when models have access to search tools, they may still hallucinate citations in contexts where retrieval fails silently or returns incomplete results.

Why AI Hallucinated Citations Are Dangerous

Academic Integrity at Scale

When a fabricated reference enters the scholarly record, it creates a cascade of problems. Other researchers may cite the same non-existent source, compounding the error. Literature reviews become unreliable. Meta-analyses may include phantom data points.

Student Assessment Becomes Unreliable

For educators, the challenge is immediate. A student paper with 30 references may contain five or six that are entirely fabricated. Traditional plagiarism detection tools like Turnitin are designed to catch copied text, not invented sources. A fabricated citation is original text — it just happens to describe a paper that does not exist.

Professional Consequences

For researchers submitting to peer-reviewed journals, including a hallucinated citation can result in desk rejection, reputational damage, or retraction of published work. The consequences extend beyond the individual paper to the credibility of the researcher's entire body of work.

How to Detect AI Hallucinated Citations

Manual Verification Methods

The most direct approach is to check each citation individually:

1. Search the DOI: Enter the DOI into doi.org or Crossref. If it resolves to nothing, the citation may be fabricated
2. Check the journal: Verify that the journal exists and has published in the volume and year cited
3. Search the authors: Confirm that the named authors have published on the cited topic
4. Look for the title: Search the exact title in Google Scholar, PubMed, or OpenAlex

This method works but does not scale. A 40-reference bibliography takes significant time to verify manually.

Automated Detection

Automated tools can verify citations against multiple databases simultaneously. The key databases for verification include:

• Crossref: The largest DOI registration agency, covering over 160 million metadata records
• OpenAlex: An open catalog of the global research system with over 250 million works
• PubMed: The primary database for biomedical and life sciences literature, with over 36 million citations
• arXiv: The preprint server for physics, mathematics, computer science, and related fields, hosting over 2.4 million papers

A citation that cannot be found in any of these databases — especially when combined with metadata inconsistencies like mismatched volume numbers or incorrect page ranges — has a high probability of being fabricated.

What to Look For

Common signals of a hallucinated citation include:

• The DOI format is valid but does not resolve to any publication
• The journal exists but did not publish in the cited volume or year
• The authors are real researchers but have no publication matching the cited title
• The title combines real concepts in a plausible but non-existent configuration
• Page numbers or volume numbers do not match any issue of the journal

Building a Verification Workflow

For Students

Before submitting any paper, run your reference list through a verification check. This is particularly important if you used AI assistance during any part of the writing or research process. Even if you did not ask the AI for citations directly, auto-complete features and writing suggestions may introduce references that need verification.

For Educators

Consider adding a citation verification requirement to your syllabus. When students know their references will be checked, they are more likely to verify them proactively. A simple policy addition can shift the behavior of an entire class:

For Researchers

Integrate citation verification into your manuscript preparation workflow. Check references after drafting but before submission. This catches not only potential AI hallucinations but also common errors like transposed digits in DOIs, incorrect publication years, or references to retracted papers.

How TrustCite Addresses This Problem

TrustCite was built specifically to solve the citation verification problem at scale. Paste your reference list, and TrustCite cross-checks every citation against Crossref, OpenAlex, PubMed, and arXiv simultaneously.

Each reference receives a verification status:

• Verified: The citation matches a real publication with correct metadata
• Minor Error: The publication exists but has metadata discrepancies (wrong volume, page numbers, etc.)
• Major Error: Significant issues detected, such as retracted papers or severely mismatched metadata
• Not Found: No matching publication found in any indexed database — a strong indicator of fabrication

After verification, you can generate a verification certificate — a digitally signed document that proves your citations were checked. This certificate can be submitted alongside your paper as proof of citation integrity.

Moving Forward

AI hallucinated citations are not going away. As language models become more sophisticated, the fabricated references they generate will become harder to distinguish from real ones. The response cannot be to stop using AI tools — they provide genuine value in research and writing. The response must be to verify.

Every citation in every paper should be checked against real databases before submission. The tools to do this exist today. The question is whether the academic community will adopt them before hallucinated citations become endemic in the scholarly record.

Sources

1. Walters, W.H. & Wilder, E.I. (2023). "Fabrication and errors in the bibliographic citations generated by ChatGPT." Scientific Reports, 13, 14045. doi.org/10.1038/s41598-023-41032-5
2. Alkaissi, H. & McFarlane, S.I. (2023). "Artificial Hallucinations in ChatGPT: Implications in Scientific Writing." Cureus, 15(2), e35179. doi.org/10.7759/cureus.35179
3. Májovský, M. et al. (2023). "Artificial Intelligence Can Generate Fraudulent but Authentic-Looking Scientific Medical Articles: Pandora's Box Has Been Opened." Journal of Medical Internet Research, 25, e46924. doi.org/10.2196/46924
4. Crossref. "Crossref Reports." crossref.org/documentation/reports
5. OpenAlex. "About OpenAlex." docs.openalex.org
6. PubMed. "About PubMed." pubmed.ncbi.nlm.nih.gov/about
7. arXiv. "About arXiv." arxiv.org/about

---

Ready to check your citations? Try TrustCite for free — verify your entire reference list in minutes.