Free AI Citation Checker Catch Hallucinated Citations Before Submission
A hallucinated citation is a reference to a paper that was never published, produced by a model that generated text shaped like a citation rather than looking anything up. It is hard to catch by reading, because every part of it looks right. Paste a reference list and each entry is resolved against Crossref, OpenAlex and Semantic Scholar, the registries holding what publishers and indexers actually deposited. Every entry gets a plain verdict, the real record when one exists, and a corrected citation ready to copy in seven styles.
and other trusted open scholarly databases
84,000+
International Researchers
Average Rating
Trusted by researchers publishing in
How to check a reference list for hallucinated citations
Add your reference list
One entry per line, a numbered list or a full pasted bibliography all parse the same way; wrapped lines and hanging indents are handled automatically.
Each entry gets resolved against registries
A DOI, where present, is resolved first. Everything else is matched by title, author and year across Crossref, OpenAlex and Semantic Scholar.
Read what each record actually says
Every entry comes back labeled, with the real record shown when one exists and a corrected citation ready in the style you need.
How this checker verifies a reference list
Grounded in registry records
Every verdict traces to an actual entry in Crossref, OpenAlex or Semantic Scholar, the community-run indexes scholarly search itself is built on.
Tuned to how AI references fail
A fabricated reference usually fails in one specific way: flawless formatting wrapped around facts that do not hold up. The checker is built to catch that pattern, above all a real DOI attached to the wrong paper.
DOIs actually resolved
Each DOI is followed to the record it points to. When that record describes a different paper than the entry claims, the checker shows you exactly what the DOI really identifies.
A corrected citation on request
Once a record is confirmed, it is reformatted into a clean reference in APA, MLA, Chicago, Harvard, IEEE or Vancouver, with italics that survive pasting into Word or Google Docs.
What a hallucinated citation check can confirm, and what it cannot
A citation makes a claim about the outside world: that a specific person wrote a specific paper, published in a specific place, in a specific year. A language model producing that citation has not checked any of it. It has generated text shaped like a reference, because that is what references look like in its training data, and the text can turn out entirely accurate, entirely invented, or a blend of several real sources stitched into one entry that describes nothing ever published.
Resolving the claim means going to the record itself rather than reading the citation more carefully. A DOI is a pointer maintained by Crossref, the agency publishers deposit metadata with directly, and following that pointer returns the actual title, author list and journal attached to it, whichever those turn out to be. Semantic Scholar and OpenAlex extend the same check to entries carrying no DOI at all, matching on title, author and year against records assembled from millions of publisher and repository sources.
A verified entry has been checked against a publisher's own deposited record, as strong a confirmation as an automated lookup can give. An entry the registries cannot find gets reported as unfound and left there, because thin registry coverage of books, theses, preprints and non-English venues is common on its own and says little about authenticity. A record that actively contradicts the entry, wrong journal, wrong year, wrong author, is the stronger evidence of fabrication, and the checker distinguishes the two outcomes rather than folding them together.
Confirming that a source exists is a narrower task than confirming what it says, and this checker only does the first. Formatting a confirmed record into a clean reference borrows the same engine as the citation generator, and citing an AI conversation itself, rather than checking one, is handled by the AI citation generator. Once the reference list is settled, the TextPulse humanizer takes on the prose surrounding it.
How common is a hallucinated citation?
Common enough to be measured at the scale of the whole literature, and rising. Two audits published within a day of each other in May 2026 put numbers on it for the first time, and both found the same direction of travel.
Papers carrying at least one fabricated reference
Per 10,000 papers, PubMed Central Open Access, January 2023 to February 2026
first seven weeks
| Period | Rate | Per 10,000 papers |
|---|---|---|
| 2023 | 1 in 2,828 | 3.5 |
| 2025 | 1 in 458 | 21.8 |
| 2026, first seven weeks | 1 in 277 | 36.1 |
The chart is drawn from an audit published in The Lancet, which screened references across roughly 2.5 million biomedical papers on PubMed Central Open Access. Of the references it could evaluate, several thousand pointed to work that does not exist, spread across more than 2,800 published papers. The rate roughly tripled between 2023 and 2025, then rose again in the opening weeks of 2026.
A companion study, Zhao and colleagues on non-existent citations, widened the net to 111 million references across arXiv, bioRxiv, SSRN and PubMed Central, and put a conservative floor of 146,932 hallucinated citations on 2025 alone. Its finding about who is affected is the more useful one for a working researcher: fabricated references cluster in fields with fast AI uptake and among small or early-career author teams, which is to say among people with the least institutional support for catching them.
Why a fabricated reference is hard to spot by reading
A hallucinated citation is not a garbled one. It is generated by the same process that writes fluent prose, so it arrives with every part a real citation has: plausible authors who publish in that field, a title that fits the argument it is attached to, a real journal, a volume and page range in the right shape, and a DOI in valid format. Nothing about it reads as wrong on the page.
The most common failure is subtler still. Rather than inventing a paper outright, a model blends fragments of several real ones into a single entry that corresponds to none of them: this author, that title, a third paper's DOI. Reading more carefully cannot separate that from a genuine reference, because every individual component is real. Only resolving the entry against the record can.
What a citation checker can and cannot do
A 2026 evaluation of five hallucinated-citation detectors, Badalova and Mayr, concluded that such tools give useful early warnings but remain constrained, and named four limits that apply to every checker in this category, including this one.
- Reference extraction. Pulling entries cleanly out of a document is its own problem. This checker takes a pasted list and splits it on entry boundaries rather than parsing a PDF, which removes that failure at the cost of asking you to paste the list.
- Incomplete metadata. An entry with no DOI, no year or a truncated title gives a registry little to match on, and a thin match is reported as one rather than promoted to a verdict.
- Database coverage. Books, chapters, theses, proceedings, very recent work and non-English venues are thinly indexed everywhere. This is why an unfound entry is reported as unfound and not as fabricated: the two are different findings and only one of them is evidence.
- Inconsistent results. Registries disagree with each other. Checking against Crossref, OpenAlex and Semantic Scholar together rather than any one of them is what narrows that, and a disagreement is surfaced instead of being resolved silently.
The practical consequence is worth stating plainly. A checker of this kind is strongest at the thing that matters most, catching a DOI that resolves to a different paper than the entry claims, and weakest on material the open registries do not hold. It is a first pass that finds the entries worth your attention, not a clearance certificate for a bibliography.
Who checks references with this tool
Writers using AI-assisted drafting
Every reference a chatbot supplied gets checked against real registries before a marker or reviewer runs the same check.
Students facing a submission deadline
One pass over the bibliography before the deadline, with any mismatched entry repaired into a corrected citation in the required style.
Faculty and graduate researchers
Reference integrity confirmed ahead of submission, including the DOI-mismatch pattern, before an editor or reviewer finds it first.
Editors and instructors reviewing submissions
A submitted reference list checked against real registry records in minutes, with verdicts built to inform rather than accuse.
The sources check out.
The sentences around them still need a pass.
The TextPulse humanizer rewrites phrasing and rhythm through the rest of the manuscript while every confirmed citation stays untouched, with each change shown as a tracked edit you approve line by line.