App Features
AI ProofreaderText HumanizerParaphrasing ToolAI SummarizerAI TranslatorFree ToolsFAQsUser GuideBlogPricingContactSign up - It's freeSign inBy signing up, you agree to our Terms and Conditions and Privacy Policy.
AI writing tools sometimes invent references that look completely real, the failure now known as a hallucinated citation. Paste your reference list and this checker verifies every entry against trusted open scientific databases like Crossref and Semantic Scholar, the databases behind scholarly search, covering hundreds of millions of records. Each reference gets a clear verdict, the real record when one exists, and a corrected citation you can copy in seven styles.
and other trusted open scientific databases
Trusted by researchers publishing in
Drop in the references an AI gave you, or any list you want to double-check. One reference per line, a numbered list or a pasted bibliography all work; the checker splits it into entries automatically.
Each entry is looked up in trusted open scientific databases like Crossref, the official DOI registry, and Semantic Scholar. DOIs are resolved first, then titles, authors and years are matched against real records.
Verified entries are done. A DOI pointing at the wrong paper is flagged with the paper it actually belongs to. Anything unverified is marked for a manual check, and every matched record comes with a corrected citation to copy.
Verification runs against trusted open scientific databases like Crossref and Semantic Scholar: hundreds of millions of scholarly records maintained for and by the academic community. Every verdict is grounded in a real lookup, never in a model's opinion.
Language models produce references that are perfectly formatted and confidently wrong. The checker is built around their failure patterns, including the most telling one: a real DOI attached to the wrong paper.
Every DOI in your list is resolved, and when it points to a different work than the entry describes, you see exactly what it really belongs to. A formatted-looking DOI is no longer evidence of anything.
When a real record is found, the checker formats it into a clean reference in APA, MLA, Chicago, Harvard, IEEE or Vancouver, with italics that survive pasting into Word or Google Docs.
A language model asked for sources does something subtly different from looking them up: it writes text that resembles the sources it saw during training. The result is a reference with everything in the right place, a plausible author team, a title that fits the topic, a journal that exists, a DOI in the correct format. Some of those references are real. Others are composites, stitched from fragments of many real papers into an entry that matches none of them. Both kinds look identical on the page, which is why every reference an AI suggests needs to be checked against a real database before it enters your manuscript.
That check is what this tool automates. Each entry in your list is verified against the most trusted open databases in scholarly publishing: Crossref, where publishers themselves register the metadata behind 160+ million DOIs; Semantic Scholar, the Allen Institute for AI's index of 200+ million papers; and other open scholarly indexes that extend coverage further still. A DOI in your entry is resolved first, because a DOI that points to a different paper than the one described is the single most reliable signature of an AI-invented reference. Entries without a usable DOI are matched by title, first author and year instead.
The verdicts are deliberately honest about what a database lookup can and cannot prove. A verified entry is confirmed against the publisher's own metadata. But an entry that is not found is reported as exactly that, not as fabricated: citation databases under-represent books, chapters, theses, preprints and non-English venues, so absence is a reason to check manually, never a verdict of guilt. Our hallucinated-citation audit walks through that manual check step by step, and it is the companion piece to this tool.
Two things this checker is not. It is not an AI detector: it makes no claim about how your text was written, and it judges only whether each reference corresponds to a findable publication. And it is not a citation formatter, although it borrows one: when a real record is found, the corrected entry is produced by the same engine as our citation generator, and if you need to cite the AI conversation itself, the AI citation generator covers ChatGPT, Claude, Gemini and the rest with the official style patterns.
Four verdicts, each grounded in what the databases actually returned. The checker tells you what it found, and just as plainly what it could not.
The entry's DOI resolves to a real record, and the record's title, first author and year match what you pasted. The strongest result a citation can get.
No usable DOI in the entry, but a real record matching its title, author and year exists in the databases. Compare the details side by side, then copy the corrected version if anything differs.
The DOI resolves to a different paper than the entry describes, or does not resolve at all. This is the classic signature of an AI-invented reference; the checker shows what the DOI really points to.
Not found in the databases we checked. That is not proof the reference is invented: books, chapters, preprints and very recent or non-English work are often absent. Locate the source yourself before keeping it.
Common enough that the question now has a published answer. A Columbia University team led by Maxim Topaz screened 2.5 million biomedical papers in PubMed Central for references that cannot be found in any major scholarly database. In the first seven weeks of 2026, one paper in every 277 carried at least one such reference. In 2023 the figure had been one in 2,828, a twelve-fold climb in two years that tracks the adoption of AI drafting tools almost exactly.
| Publication window | Papers citing at least one unverifiable reference |
|---|---|
| 2023 | 1 in 2,828 |
| First seven weeks of 2026 | 1 in 277 |
Source: a Columbia University audit of 2.5 million PubMed Central papers, published in The Lancet, 2026.
Peer review is not the safety net it appears to be here. When GPTZero screened the 4,841 papers accepted at NeurIPS 2025, it confirmed at least 100 hallucinated citations across 53 of them, in work that three to five expert reviewers had already approved. The failure is also harder to see than most authors expect: in a Deakin University test of GPT-4o-generated references, 56 percent were fabricated or wrong, and 64 percent of the fabricated ones arrived with a genuine DOI that resolves to an unrelated paper. That is exactly the case a visual check clears and a database lookup flags, and it is why this checker resolves every DOI rather than trusting the formatting.
Journals increasingly treat an invented reference as an integrity question rather than a typo, and arXiv can suspend submission privileges for a year over a single unchecked one. Running your list through this checker takes minutes; the full pre-submission routine, including the manual escalation steps, is in our hallucinated-citation audit.
ProofreaderPro proofreads your full manuscript with tracked changes and never touches your citations: grammar, clarity and academic register, reviewable line by line. Built for researchers, free to try.
Proofread my writing freeProofread, humanize, paraphrase and translate your academic writing in one place. Free to start, no credit card required.
Get Started Free