1. Fabricated references also appear in published journal articles
A research letter by Topaz and colleagues, published in The Lancet in May 2026, audited approximately 2.47 million biomedical papers in the PubMed Central (PMC) Open Access subset, published from January 1, 2023, to February 18, 2026. Of approximately 125.6 million references, the researchers verified about 97.1 million carrying a PMID. References that could not be found in PubMed, Crossref, OpenAlex or Google Scholar were classified as fabricated. The audit identified 4,046 such references across 2,810 papers. [1][2][3]
According to the research team’s institution, Columbia University School of Nursing, the rate was approximately 4 fabricated references per 10,000 papers throughout 2023. It rose sharply from mid-2024, reaching approximately 57 per 10,000 papers in early 2026—more than 12 times the 2023 rate. [2]
The proportion of papers containing at least one fabricated reference is a different measure: approximately 1 in 2,828 papers in 2023, 1 in 458 in 2025 and 1 in 277 during the first seven weeks of 2026. The 2026 observation period ended on February 18. At the time of the audit, 98.4% of affected papers had received no publisher action. [3][2]
A July 2026 correction reports changes to an author’s affiliations and the appendix. Read linked corrections alongside the original paper when interpreting the research.
Keep these limitations in mind
- Scope: The sample covered PMC Open Access papers, so it does not directly represent the entire literature, including subscription publications.
- Verification coverage: Approximately 23% of references had no PMID and were excluded. Excluded references and errors missed by the pipeline mean the findings are not an estimate of every fabricated reference in the literature.
- Relationship to AI: The audit did not establish whether AI generated each reference. The rise coincided with the spread of AI writing tools; that timing does not establish the cause.
- Classification: Failure to find a reference in four databases was used to classify it as fabricated. Some real publications outside those indexes could have been included.
The audit nevertheless shows why authors need to verify references themselves instead of relying on peer review to catch every problem. [2][3]
2. Why can conversational AI invent a paper?
Language models generate plausible next expressions from patterns learned in text. When answering without a search tool, a model may combine familiar author names, journal titles, years and topic terms into a reference-shaped sentence instead of retrieving a paper from a database.
Watch for these common patterns:
- A nonexistent title attributed to real researchers and a real journal
- A real paper with an altered title, year, volume or issue
- A DOI belonging to another paper, so the link opens but the content is wrong
- A real paper whose results or conclusions differ from the AI’s account
Tools with search features can also connect a claim to the wrong source. Whatever tool you use, you remain responsible for the references in your manuscript. ICMJE recommends reviewing and editing AI output because it can be inaccurate, incomplete or biased; human authors must ensure appropriate attribution and full citations. [4]
3. Verify each reference in five steps
Check references individually. If a search fails at one step, cross-check at the next rather than immediately declaring the paper fabricated.
Search the complete title in quotation marks
Enter the full title in double quotation marks in Google Scholar or PubMed. Google Scholar’s help recommends this when looking for a particular paper. [5] In PubMed, quotes are ignored if the phrase is absent from its phrase index; check the message above the results to see how the search was handled. [6]
No results means further checking is needed. The AI may have altered a real title, or the publication may not be indexed in that database. Continue with the DOI and other bibliographic details.
Open the DOI and compare the bibliographic details
If a DOI is supplied, type
https://doi.org/followed by the DOI in your browser. A DOI link resolves to current information about the identified object. [7] For example, the Lancet audit discussed here is https://doi.org/10.1016/S0140-6736(26)00603-3.An opening link is not enough. Compare the title, authors, journal and year on the destination page with every detail the AI supplied. If the DOI does not resolve or remains uncertain, search Crossref by title or DOI. [8]
Cross-check the author, journal and year
PubMed’s Single Citation Matcher can locate papers from partial information such as journal, year, volume, first page and author. [9] Check whether the author published a relevant paper in that journal and year, and compare volume, issue and page numbers. If you locate the real paper, correct the reference from its actual record rather than retaining the AI’s version.
Find the passage that supports your sentence in the original paper
Existence and support are separate questions. Read the original paper’s population, design, results and limitations rather than citing an abstract or AI summary alone. Decide whether your claim stays within the evidence. Marking the supporting passage also makes it easier to answer a reviewer’s question later. See section 4 for how to judge and record claim support.
Check for retractions and corrections
A paper can exist and appear relevant yet have been retracted. PubMed links original articles to retraction notices based on information published by journals; inspect any linked retraction or correction notices. [10] You can also consult the Retraction Watch database made publicly available by Crossref. [11]
Apply the same checks when copying a reference from another paper’s bibliography. Most papers containing fabricated references in the audit had received no publisher action at that point. [2]
4. The paper is real—but does it support your claim?
Jump to the worked example with a real source table →
A reference can exist and have correct bibliographic details and still be the wrong citation. The finding may be weaker than you wrote, may come from a different comparison, or may not appear in the source at all. Once you have confirmed that a reference exists (section 3), compare each cited claim with its evidence. You can do this in four steps without any tool.
Isolate the claim in the cited sentence
Mark the population, comparison, outcome, number and direction (increase, decrease or association). If a sentence cites several sources, note which part each source is meant to support.
Find the passage in the original paper
Look in the Results text and tables rather than stopping at the abstract, and note the page or table number. If you cannot access the full text, record the reference as “could not check.”
Classify the match
(1) Supports the claim as written; (2) supports only a narrower version, such as a subgroup result or an association rather than an effect; (3) the information is not in the source; (4) the source contradicts it. A matching number is not enough: check that the comparison, unit and denominator are the same.
Revise according to the verdict
For (2), narrow the wording to match the source. For (3) and (4), replace the source or remove the claim. Keep “could not check” references on your pre-submission list until you have read the original.
Record your verdicts
A simple log makes it easier to divide the checking among coauthors and to answer a reviewer who asks where the evidence is.
| Claim (short) | Reference and location | Verdict | Action |
|---|---|---|---|
| “Complications were twice as common in group A” | Ref. 12 · Table 2 | (2) Narrower: patients aged 65 or older only | State the subgroup |
| “Drug B reduced mortality” | Ref. 7 · Results text | (3) Not in the source: only readmission was reported | Revise the sentence or replace the source |
| “Marker C was associated with prognosis” | Ref. 21 · full text unavailable | Could not check | Keep it on the list until you obtain the original |
When moving verified evidence into the Introduction, distinguish what each paper answered from what remains unresolved. Continue with the evidence-comparison table and gap-to-objective example in the Introduction guide.
Practice: the DOI matches, but what does 55% mean?
The incorrect sentence below was invented for this exercise. Open Methods and Table 3 in the linked real paper and check what the percentage counts, as well as the number itself.
Before · incorrect claim
“Today, 55% of AI-written medical papers are fabricated.”
Where to check the source
Walters and Wilder (2023), Methods and Table 3: All works row, GPT-3.5 column. The 55% refers to fabricated references among the 222 works cited by GPT-3.5 in this study.
Open the real paper and Table 3 ↗ · DOI: 10.1038/s41598-023-41032-5
Judgment · supports a narrower claim
A percentage of references became a percentage of papers. This result also does not justify extending a particular model and task to all current AI or all medical papers.
After · our paraphrase
“In Walters and Wilder’s 2023 study, 55% of the 222 references in GPT-3.5-generated documents were classified as fabricated.”
Source checked November 8, 2026. This is not an evaluation of current models or Paper Buddy’s error rate.
Apply this to one citation in your manuscript
My sentence and citation number: ______
Paper title and DOI / source page, table and row: ______
Supported scope and anything not verified: ______
Judgment and revised sentence: ______
Checker and date / remaining action: ______Copy these five lines into your document and complete them. Leave “not verified” if you cannot obtain the source. Recheck the revised claim and make sure its citation still points to the intended reference.
Compare my sentence with the source passages and table I checked below. Identify differences in population, unit of analysis, denominator, time and claim strength. Suggest a revision within the supplied evidence. Mark missing information as not verified; do not invent references or numbers.
Recheck citations after AI editing or translation
Even a citation you checked earlier needs another comparison with the source after AI rewrites, shortens or translates the sentence. The reference number can stay the same while the population, conditions or strength of the claim changes.
On small screens, scroll the table horizontally to compare all three columns.
| Wording checked against the source | Revision to recheck | What to check |
|---|---|---|
| “A was associated with B in patients aged 65 or older.” | “A was associated with B in patients.” | Has removing the age condition broadened the population? |
| “A was associated with B.” | “A caused B.” | Has an association become a causal claim? |
| “A was associated with B.[12]” | “A was associated with B and also improved C.[12]” | Does the existing reference support the added claim? |
If sentences were combined or reordered, check which claim each citation number supports. Apply the four steps above to the revised sentence and record the date of the new check and the source location in the log’s action column.
For a review, inspect the whole reference list too
After checking each source, look across the list for these issues.
- Duplicate records for one paper: Entries with different title or author formatting may share a DOI. Distinguish duplicate records from citing the same paper in several sentences.
- A clear connection to the cited sentence: Sources from other fields can provide relevant evidence. Check that you can explain each citation rather than excluding it because of its field.
- Coverage that matches the review’s scope: If a few papers carry much of the argument, check whether the evidence reflects the population, period and perspectives the review promises to cover.
If you use an AI citation checker: SciSpace Citation Verifier as an example
AI tools that compare claims with cited sources are also appearing. SciSpace released Citation Verifier on October 7, 2026. According to SciSpace, References mode checks whether cited papers exist and whether publication details match, and flags retractions or corrections that need attention. Full mode compares cited claims with the source material SciSpace can retrieve and read, and shows whether the source supports the statement, supports a narrower version, does not contain the claimed information or contradicts it. When a readable source is unavailable, the claim is marked “Could not be checked.” SciSpace notes that a claim marked “not supported” is not necessarily false; the source reviewed does not establish it as written. The announcement does not state pricing or whether the free plan includes the feature, and we have not run it. [12]
Whatever tool you use, open the passage behind each verdict yourself and check that any suggested rewrite stays within the source. Tools that check only existence, bibliographic details and retractions can tell you that a paper is real without telling you whether it supports your sentence. Before uploading an unpublished manuscript, check the service’s data handling and your institution’s rules.
In Paper Buddy, you can show Paper Chat a passage selected in a PDF and ask how far it supports a sentence in your manuscript. Use the answer as a guide, but make the four-way judgment yourself after reading the original.
5. Find the literature, then ask AI to work with the sources you read
When you need evidence, search indexed literature in databases such as PubMed or Google Scholar, or use tools that search those records. Then ask AI to summarize or check wording against the sources you found, rather than to invent a reference list without source material.
| Approach | Example request | Reason |
|---|---|---|
| Avoid | “Find three papers published since 2024 supporting this claim and format the references.” | Without source material, a plausible list may include nonexistent papers or papers about something else. |
| Prefer | “Use the Results of this paper I found to check whether my Discussion sentence exceeds the evidence. Flag anything not established in the original text.” | The source is specified, so you can compare the answer directly with the relevant passage. |
In Paper Buddy, you can collect literature through PubMed and Semantic Scholar search or import PDF, BibTeX, RIS and NBIB files. Ask Paper Chat about a selected PDF passage, then connect verified sources to manuscript citations. Imported metadata can still be incomplete; compare it with the original using the five steps above. See literature search, PDF reading and citation features →
Before uploading a PDF or full text to an AI tool, check its permitted uses, your institution’s rules and the tool’s data handling policy. Take particular care with unpublished manuscripts and patient information.
If you used AI for literature exploration or writing, check the target journal’s disclosure requirements. ICMJE states that AI-generated material must not be cited as a primary source, and that nondisclosure may require corrective action or be considered misconduct in some circumstances. Check the specific disclosure requirements of your target journal. [4] Read the AI disclosure guide → See disclosure wording in the submission guide →
6. Before-submission reference checklist
- Have you confirmed that every reference exists through a title search or DOI?
- Do the DOI destination’s title, authors, journal and year match the reference?
- Have you found a supporting passage in the original for each cited claim?
- Have you narrowed sentences that the source supports only in part, and resolved every reference you could not check?
- Have you checked and addressed linked retractions and corrections?
- Have you removed or verified any AI-provided wording adopted without reading the original?
- Do in-text citations and bibliography entries correspond to each other?
- Have you compared AI-edited, shortened or translated sentences with the original sources again after revision?
- For a review, have you checked the whole list for duplicate records and off-topic citations, and can you explain why each source is cited?
Frequently asked questions
Is a paper necessarily fake if Google Scholar cannot find it?
No. Google Scholar does not guarantee continuous, complete coverage of any source, and a slightly altered title can prevent a match. Cross-check the DOI, author, journal and year. If you still cannot verify the paper, omit it or replace it with a source you have found and checked. [5]
Does an opening DOI link prove the reference is correct?
It only shows that the DOI resolves to an object. It might belong to a different paper. Compare the destination’s title, authors, journal and year with the reference you were given. [7]
Do I need to recheck references if AI only polished the wording?
Yes. Even when the reference stays the same, changes to the population, conditions or strength of a sentence, or an added claim, can leave the revised wording unsupported. Compare the sentence and citation placement with the original source again. Follow your target journal’s AI disclosure requirements. [4]
The reference is real, but how do I check that it supports my claim?
Mark the population, comparison, outcome and number in the cited sentence, then find the matching passage in the Results text or tables of the original paper. Classify it as supporting the claim as written, supporting only a narrower version, not in the source, or contradicting it. A matching number can still refer to a different comparison or unit. Narrow your wording if the support is partial; replace the source or remove the claim if the information is absent or contradicted.
If I use a citation checker such as SciSpace Citation Verifier, can I skip checking myself?
No. SciSpace says Full mode compares claims with the source material it can retrieve and read, and marks claims without a readable source as “Could not be checked.” Treat a verdict as a pointer to where to look: open the passage behind it, and check that any suggested rewrite stays within the source before you accept it. ICMJE also recommends that authors review and edit AI-generated output. [12][4]
Sources
Sources checked on October 19, 2026; the SciSpace announcement was checked on October 21, 2026. The retraction notice and ICMJE guidance were checked on October 29, 2026. The year-specific proportions and increase in rate are reported in the research team’s institutional release and Retraction Watch coverage.
- Topaz M et al. Fabricated citations: an audit across 2·5 million biomedical papers. Lancet 2026;407:1779–81
- Columbia Nursing: announcement of the fabricated-citation audit (May 2026)
- Retraction Watch: year-specific proportions and limitations of the audit
- ICMJE: use of AI by authors
- Google Scholar search help: title searches and coverage
- PubMed User Guide: searching for a phrase
- DOI Foundation: resolving a DOI
- Crossref metadata search
- PubMed Single Citation Matcher
- NLM: links to corrections and retractions in PubMed
- Crossref: Retraction Watch database documentation
- SciSpace: Introducing SciSpace Citation Verifier (October 7, 2026)
- Springer Nature: Discover Artificial Intelligence retraction notice (October 3, 2026)