Fabrication and Errors in the Bibliographic Citations Generated by ChatGPT
What Walters and Wilder's 2023 Scientific Reports study did, every headline figure as published, what it cannot tell you, and how to use it when checking your own references.
"Fabrication and errors in the bibliographic citations generated by ChatGPT" is a 2023 paper in Scientific Reports by William H. Walters and Esther Isabelle Wilder. They asked ChatGPT for 84 short literature reviews and checked all 636 citations. They found 55 percent of GPT-3.5 citations and 18 percent of GPT-4 citations were fabricated, and many real ones still carried errors.
This guide reports the method and the figures exactly as the paper publishes them, with the paper itself linked through its DOI. It then explains what the study can and cannot tell you, and how to use its categories when you check references in your own thesis.
Who wrote the paper and where it appeared
The paper was published in Scientific Reports, volume 13, article number 14045, on 7 September 2023. It was received on 11 July 2023 and accepted on 21 August 2023, and it is open access under a CC BY 4.0 licence, according to the Crossref record for the DOI. That record also gives the title, authors, volume and article number.
William H. Walters is listed at the Mary Alice & Tom O'Malley Library, Manhattan College, Riverdale, New York. Esther Isabelle Wilder is listed at the Department of Sociology, Lehman College, The City University of New York, and at the Doctoral Program in Sociology, CUNY Graduate Center. Walters is the corresponding author.
The record lists it as volume 13, issue 1. APA Style's journal article guidance says to include the issue number and to write "Article" before an article number, so the APA entry reads:
Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13(1), Article 14045. https://doi.org/10.1038/s41598-023-41032-5
How the study worked
The authors used two versions of ChatGPT, GPT-3.5 and GPT-4, to write short literature reviews on the same 42 topics. That gave 84 papers and 636 citations, which they then checked one by one against databases and websites.
The prompt and the topics
Each topic started a new chat. The prompt began "I want you to act as an academic researcher" and asked for a paper of about 2,000 words "with parenthetical citations and a bibliography" of at least five scholarly sources. When a reply stopped early, the authors typed "Please continue", and they ignored any text after the first bibliography.
The topics were broad and mostly social, political or environmental. Examples named in the paper include the health effects of e-cigarettes, the original purpose of Stonehenge and the effects of Brexit on the U.K. economy. All texts were generated in the first week of April 2023.
How each citation was checked
The authors searched Google, Google Scholar, Amazon, the Directory of Open Access Journals, PubMed, Scopus, WorldCat and publishers' and journals' websites. For works that still looked fabricated, they browsed the journal's volume and issue and searched the publisher's site as a final check. All the data are provided in the paper's supplementary appendices.
How the authors defined each problem
The study sorts every citation with three definitions. They matter, because other studies define "fake" differently.
| Term in the paper | What it means |
|---|---|
| Real | An actual work that matches the citation closely on both title and authors |
| Fabricated | A work that has never been published, presented, posted or otherwise publicly disseminated |
| Substantive error | A real work cited with wrong authorship, title, journal, publisher, volume, pagination or similar details, or an omission that makes it hard to find |
| Formatting error | A breach of APA Style, such as title capitalisation or punctuation, that does not stop you finding the work |
What "fabrication and errors in the bibliographic citations generated by ChatGPT" found
The short version: GPT-4 did much better than GPT-3.5, and neither was safe to trust. The authors' own summary is that GPT-4 "is a major improvement over GPT-3.5" but "problems remain".
All figures below come from the paper's tables. We read them in the full text on PubMed Central, which carries the same article.
| Measure | GPT-3.5 | GPT-4 |
|---|---|---|
| Papers written | 42 | 42 |
| Works cited | 222 | 414 |
| Mean works cited per paper | 5.3 | 9.9 |
| Cited works that were fabricated | 55% | 18% |
| Real works with at least one substantive error | 43% (of 101) | 24% (of 340) |
| All citations with at least one formatting error | 45% | 40% |
Fabrication by type of source
The paper splits fabrication by what was cited. Book chapters stand out: 70 percent of cited chapters were fabricated for both versions. The chapter counts are small (10 and 23), so treat that row with care.
| Type of work (number cited) | GPT-3.5 fabricated | GPT-4 fabricated |
|---|---|---|
| Journal articles (128 and 255) | 73% | 18% |
| Books (78 and 126) | 23% | 8% |
| Book chapters (10 and 23) | 70% | 70% |
| Websites (6 and 10) | 50% | 10% |
Errors in the real citations
When the cited work did exist, the details were still wrong in many entries. For GPT-3.5, 22 percent of real works had the wrong date and 34 percent of articles and chapters had wrong volume, issue or page numbers. For GPT-4 those figures were 16 percent and 13 percent.
Wrong author names appeared in 14 percent of real GPT-3.5 citations and 6 percent of real GPT-4 citations. Wrong journal titles appeared in 14 percent and 4 percent of articles. Numbers, in other words, broke more often than names.
Formatting and links
Every citation was in APA Style, and improper title capitalisation was the most common formatting error (44 percent and 39 percent). Links were rare: 5 percent of real GPT-3.5 citations and 8 percent of real GPT-4 citations included one. Of those links, 60 percent and 70 percent led to the cited work.
The finding that matters most for a thesis: fake references borrow real parts
The most useful detail for anyone checking a reference list is how the fabricated citations were built. The paper reports that most fabricated article, book and website citations named real journals, publishers or organisations.
Table 3 makes this concrete. Among fabricated articles, only 2 percent named a journal that does not exist, for both versions. Among fabricated book chapters, the parent book was itself invented in 57 percent of GPT-3.5 cases and 19 percent of GPT-4 cases.
So a familiar journal name proves nothing. The test that works is whether a matching record exists, which is why the definition of hallucinated citations rests on that missing record.
Why the authors think fabrication happens
The authors describe ChatGPT as "fundamentally not an information-processing tool, but a language-processing tool". They suggest it may not treat a citation as a special kind of text that has to be copied exactly. They also borrow the phrase "confidently wrong" from Gravel and colleagues to describe its incorrect answers.
They note that search tools such as Google Scholar "have not had such high fabrication or hallucination rates". The mechanism is covered in more depth in our guide on why AI tools make up citations.
The errors were not limited to reference lists. Five of the papers were presented as empirical studies, "with fabricated methods and results". The authors also report at least two fabricated GPT-3.5 citations attributed to journals from publishers identified as predatory.
What the authors recommend
The paper gives three short recommendations. Users of ChatGPT "are cautioned to check the citations it generates". Instructors tasked with detecting AI-generated text, in undergraduate research papers for instance, "may find it helpful to examine the works cited in those papers".
Editors and publishers should "ensure that fabricated citations do not find their way into the scholarly literature". At the time of writing, the authors add, "no publicly available AI text detector checks for fabricated citations".
How it compares with other studies
Other studies report different rates, partly because they define a fake reference differently. Comparing headline percentages without the definitions gives a false picture.
| Study | What was tested | Headline figure | How references were judged |
|---|---|---|---|
| Walters and Wilder (2023) | 636 citations in 84 short reviews, GPT-3.5 and GPT-4 | 55% and 18% fabricated | Never published or publicly disseminated in any form |
| Bhattacharyya et al. (2023) | 115 references in 30 medical papers, ChatGPT-3.5 | 47% fabricated, 7% real and accurate | Searched in Medline, Google Scholar and DOAJ |
| Chelli et al. (2024) | References for 11 systematic reviews, GPT-3.5, GPT-4 and Bard | 39.6%, 28.6% and 91.4% hallucinated | Any two of title, first author and year wrong |
Walters and Wilder also summarise six earlier studies in their introduction. Across those, 51 percent of 732 citations were fabricated.
What the study does not tell you
The paper describes two versions of ChatGPT in the first week of April 2023, so it does not measure any tool released since. Its 42 topics were broad, mostly social, political and environmental, and pitched at the level expected in first-year composition courses at U.S. universities.
A doctoral literature review on a narrow topic is a different task. The headline rates are also per citation, not per paper. Do not quote it as "half of AI papers have fake references", because it does not say that.
How to use these findings on your own references
Use the paper's categories as a sorting sheet. For every reference that came from, or passed through, a chatbot, decide which of four boxes it belongs in before it stays in your thesis.
| Box | Test | What to do |
|---|---|---|
| Real and correct | A record exists and every field matches | Keep it, and confirm it supports your sentence |
| Real with a substantive error | A record exists but the year, volume, pages, journal or authors differ | Correct every field from the publisher's record |
| Real with a formatting error | Details are right but the style is off, such as title capitals | Fix it during your style pass |
| Fabricated | No record anywhere after a full search | Delete it and find a real source for the claim, or cut the claim |
Two rules follow directly from the figures. Check the numbers (year, volume, issue, pages), because they broke most often. Search hardest on book chapters, because they had the highest fabrication rate.
Our guide to checking if a citation is real gives the full per-reference method. It covers the DOI, Crossref, the publisher's page, a library catalogue and Retraction Watch.
What Thesisroom does with this
Thesisroom's Citation Check looks each reference up in eight registries: Crossref, OpenAlex, Semantic Scholar, PubMed, Retraction Watch, DOAJ, CORE and Unpaywall. It shows which references it could verify, which are retracted, and which need checking by hand.
Thesisroom never calls a reference fabricated. A reference it cannot find goes to you for a manual search, because a missing record can also mean a book, a report or an older paper that registries do not hold.
Frequently asked questions
Who are the authors of "Fabrication and errors in the bibliographic citations generated by ChatGPT"?
William H. Walters of the Mary Alice & Tom O'Malley Library at Manhattan College, and Esther Isabelle Wilder of Lehman College and the CUNY Graduate Center, The City University of New York. Walters is the corresponding author.
What percentage of ChatGPT citations were fake in the Walters and Wilder study?
Fabricated: 55 percent of 222 GPT-3.5 citations and 18 percent of 414 GPT-4 citations. Among real works, 43 percent of GPT-3.5 citations and 24 percent of GPT-4 citations had at least one substantive error, such as a wrong date or page numbers.
When was the Walters and Wilder paper published, and is it free to read?
It was published in Scientific Reports on 7 September 2023, after being received on 11 July and accepted on 21 August 2023. It is open access under a CC BY 4.0 licence, so you can read the full text free through its DOI.
Do the 2023 figures apply to newer chatbots?
The study cannot tell you. It tested GPT-3.5 and GPT-4 in the first week of April 2023 on 42 broad topics, so newer tools need their own testing. What still applies is the lesson: a citation is only trustworthy once you have found its record yourself.
What type of reference was most likely to be fabricated?
Book chapters. Seventy percent of cited chapters were fabricated for both GPT-3.5 and GPT-4, although the counts were small, 10 and 23 chapters. For journal articles the rate was 73 percent for GPT-3.5 and 18 percent for GPT-4.
How do I cite this paper in APA Style?
Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13(1), Article 14045. https://doi.org/10.1038/s41598-023-41032-5
Sources
- Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13(1), Article 14045
- Full text of Walters and Wilder (2023) on PubMed Central (PMC10484980)
- APA Style: journal article references
- Crossref metadata record for 10.1038/s41598-023-41032-5
- Bhattacharyya, M., Miller, V. M., Bhattacharyya, D., & Miller, L. E. (2023). High rates of fabricated and inaccurate references in ChatGPT-generated medical content. Cureus
- Chelli, M., et al. (2024). Hallucination rates and reference accuracy of ChatGPT and Bard for systematic reviews: Comparative analysis. Journal of Medical Internet Research, 26, e53164
How this guide was made: Thesisroom drafted this guide with the help of AI writing tools, then checked every fact, figure and source by hand against the references linked above.