How to Check If a Citation Is Real: A 4-Step Verification Method

Fabricated citations are no longer a rare problem

Fabricated academic citations used to be a career-ending curiosity. Now they are a bibliography-scale risk. A May 2026 audit by Topaz and colleagues in The Lancet audited millions of PubMed references. They found that one of every 277 papers indexed in PubMed in early 2026 contains a fabricated reference, up from one in every 2,828 in 2023. That is roughly a tenfold rise in three years, and the rise tracks the adoption of AI-assisted drafting tools in academic writing (Topaz et al., 2026).

Knowing how to check if a citation is real matters at every stage: reading someone else's paper, writing your own literature review, and preparing a thesis for examination. The four-step method below runs on any single citation by hand. At the end, we describe the tool that runs the same checks at bibliography scale.

Why bibliography errors happen more than before

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. The remainder were partially wrong: real authors and journals stitched to nonexistent papers, or real papers cited with wrong details. Walters and Wilder (2023) confirmed the pattern at scale: across 636 citations in 84 literature reviews, 55% of GPT-3.5 references and 18% of GPT-4 references were fabricated.

A candidate who used an AI tool to draft any part of a literature review has, on the base rates alone, a meaningful chance of at least one fabricated entry in the bibliography. The candidate does not have to know it is there for an examiner to find it. Every viva has a citations question, and a fabricated entry surfaced by the panel changes the whole conversation. This is why the Thesisroom method treats bibliography verification as a whole-degree seat, not a submission-week task.

Step 1: Confirm the DOI resolves

Every real academic paper has metadata registered in an index. For papers with DOIs (the majority of journal articles from 2000 onwards), the fastest check is a DOI lookup against Crossref, the DOI registry for scholarly literature.

Take the DOI in the citation and paste it into https://doi.org/[DOI]. A resolving URL that leads to the paper on the publisher's site is a positive signal. A 404, a "DOI not found" error, or a redirect to an unrelated page is a strong warning that the citation is either wrong or fabricated. You can also query the Crossref REST API directly at https://api.crossref.org/works/[DOI] for the raw metadata.

There is a common failure mode. AI-assisted drafts often produce DOIs that look valid but do not resolve. If the DOI does not resolve, the citation is wrong until proven otherwise. The next steps rule out honest typos and confirm fabrication.

Step 2: Match the metadata against the citation

If the DOI resolves, do not stop there. Compare every field in the returned metadata against the citation in the manuscript: authors (all of them, in order), title (word for word), journal name, year, volume, issue, and pages. Discrepancies here are common and revealing.

A mismatch in authors is the most damaging error at examination. A mismatch in year or volume usually means the source was misremembered, not invented. A perfectly matching DOI with a plausible but slightly wrong title sometimes signals a fabricated citation that happens to have collided with a real DOI, which does occur in AI-generated bibliographies.

If the paper has no DOI (some books, older papers, some grey literature), verify against multiple indexes: OpenAlex (https://openalex.org), Semantic Scholar (https://semanticscholar.org), and Google Scholar. OpenAlex is the largest open scholarly index with more than 200 million works and is a strong default (Priem et al., 2024, arXiv). Semantic Scholar is stronger for computer science, biomedicine, and citation-network questions. Google Scholar has broad coverage but lower precision.

Step 3: Check for retraction status

A real citation that has not been retracted is not the same as a real citation that has. Chelli and colleagues (2024) found that a substantial fraction of AI-generated citations reference retracted papers as if they were current, because the training data predates the retraction. Retraction Watch, now integrated into Crossref's API since September 2023, is the primary index of retracted, corrected, or expression-of-concern papers.

Search the paper title or DOI at https://retractionwatch.com or query the Crossref API for the retraction flag. If a paper you cite has been retracted, either drop the citation or, if the citation is essential to a historical claim, cite the retraction and the reason. Silent citation of retracted papers is one of the failure modes examiners are trained to look for (Park, 2003).

There is a further wrinkle. Papers get retracted after they are cited in a thesis. A citation that was fine last year may be retracted this year. This is the specific gap that no reference manager fills, because reference managers store references, they do not re-verify them. The Thesisroom living bibliography re-checks entries against Retraction Watch on an ongoing basis and surfaces new retractions as they happen.

Step 4: Verify the source supports the claim

The first three steps confirm the citation is real. The fourth step confirms the citation actually supports the sentence it sits under. This is where careful writers still fail, and where AI-assisted drafts frequently fail.

Read the abstract at minimum. If the abstract does not obviously support the sentence in your text, open the paper and find the specific passage the citation is anchored to. If that passage does not exist, the citation is wrong even if the paper is real. Bhattacharyya and colleagues (2023) note that this failure mode is nearly as common as full fabrication in AI-assisted drafts: real papers cited for claims those papers do not actually make.

This step is slow by hand, and it is the reason full-bibliography verification is unrealistic without help. A doctoral bibliography of 200 to 400 sources cannot be verified for both existence and claim-fit in a submission week. The verification has to run alongside the writing.

What Thesisroom does with this

Thesisroom's citation guard runs the first three steps automatically against Crossref, OpenAlex, Semantic Scholar, Retraction Watch, and DOAJ. Every entry in the bibliography returns as verified, flagged, or unverifiable, and every flag names the specific database check that produced it. The full method treats citation verification as a habit that starts in year one of the degree. The bibliography that protects a candidate at the viva is the same one they built during their literature review.

The tool does not verify the fourth step for you: whether a real source supports a specific claim is a reading task the writer has to do. It does surface flagged sources and offer replacements through Source Finder, so the reading is targeted at the sources most likely to have been miscited.

Frequently asked questions

How can I tell if ChatGPT gave me a fake citation?

Start with the DOI. Paste it into https://doi.org/[DOI] and see if it resolves. If it does not resolve, the citation is likely fabricated. If it does resolve, check that the authors, title, journal, and year match the citation exactly. AI-generated fake citations often have valid-looking DOIs that do not resolve, or resolve to a different paper.

Are all AI citations fake?

No, but a substantial fraction are. Walters and Wilder (2023) found that 55% of GPT-3.5 citations and 18% of GPT-4 citations across 636 references in 84 literature reviews were fabricated. Newer models fabricate less than older ones, but every AI-assisted bibliography needs verification before submission.

What is the difference between Crossref, OpenAlex, and Semantic Scholar?

Crossref is the DOI registry: the authoritative source for whether a paper has a valid DOI and what metadata is registered. OpenAlex is the largest open scholarly index and covers over 200 million works. Semantic Scholar adds AI-derived citation-network features and is strong for computer science and biomedicine. Cross-checking against more than one increases coverage.

How do I check if a paper has been retracted?

Search the paper title or DOI at Retraction Watch (https://retractionwatch.com). Retraction Watch was integrated into the Crossref API in September 2023, so a Crossref query also returns retraction status. Reference managers do not check this automatically, which is why sources verified during a literature review still need re-checking before submission.

Do I need to check every citation in my bibliography?

Yes, especially in a thesis. Examiners spot-check citations that look important or unfamiliar, and a single fabricated entry surfaced in a viva changes the tone of the whole examination (Park, 2003). Manual verification of 200-plus sources is slow. Tools that run the DOI, metadata, and retraction checks at bibliography scale exist for exactly this reason.

What should I do if a citation cannot be verified?

If the DOI does not resolve and no cross-index search finds a matching paper, remove the citation and either delete the sentence or find a real source that supports the claim. Do not leave a suspect citation in a thesis on the assumption that no examiner will check. The base rates for examiners spot-checking are high enough that the assumption fails.

One takeaway before you close this tab

If you use any AI-assisted drafting tool during your literature review, plan for verification. The DOI-resolve step catches the majority of fabricated citations in under a minute per entry. The retraction check catches the sources that have moved since you saved them. Every entry your bibliography holds should have been through both, at least once. When you want the whole bibliography checked in a single pass, and re-checked as new retractions land, Thesisroom's citation guard is where the checks live.

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.

Park, C. (2003). Levelling the Playing Field: Towards Best Practice in the Doctoral Viva. Higher Education Review, 36(1), 47-67.

Priem, J., et al. (2024). OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv.

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.