Fake citations, misquoted sources: how ProofVerity verifies references
In shortLanguage models invent references that look real: plausible authors, a real journal, a volume that exists, an article that does not. ProofVerity resolves every citation in a document against thousands of journals, books, archives and registries, and names each failure: not found, misquoted, misattributed or invented. Verified references stay visible too.
Why fabricated references are the clearest AI signal
A paraphrasing tool can hide flat prose. Nothing can make a non-existent paper exist. The reference list is where machine-written work is most checkable, and where the consequences of not checking are most public: retractions, corrections and failed theses usually start with a citation nobody opened.
The four ways a citation fails
| Status | What it means | Example from a sample report |
|---|---|---|
| Not found | No record in the indexed sources. The journal may exist; the article does not. | [7] Nakamura & Osei (2021): 0 matches in 14,212 sources; nearest match is a different 2019 paper. |
| Misquoted | The source exists but does not make the claim attributed to it. | [3] Halvorsen & Ruiz (2019): DOI resolves; the source studies daytime exposure, not nighttime mortality. |
| Misattributed | The claim exists but belongs to a different work, author or year. | A quotation traced to a later review article rather than the original study cited. |
| Verified | The source resolves and supports the claim as written. | [1] Copernicus Land Monitoring Service: dataset resolves, thresholds match. |
Verified citations stay visible in the report, so a reader can see what was checked and not only what failed.
What ProofVerity checks against
- Journal indexes and DOI registries.
- Preprint servers and institutional repositories.
- Library catalogues and ISBN registries for books and chapters.
- News archives, for quotations attributed to reporting.
- Court, patent and government registries, for legal and policy citations.
- The source text itself, when it is open or licensed, to confirm that the cited sentence says what the document claims.
How a check runs
- Parse each reference and each in-text marker, whatever the citation style.
- Resolve identifiers first (DOI, ISBN, URL), then metadata (authors, title, year, venue).
- Fetch or match the source text where available.
- Compare the sentence that cites the source with what the source says, including any number attributed to it.
- Report a status with the evidence: matches found, nearest match when none, and the exact discrepancy when the source disagrees.
Numbers are citations too
A figure attributed to a source is checked against the source. In the sample report, "38% of the EU population" fails twice: the reference does not exist, and the statistic itself differs from what Eurostat publishes. Both findings appear on the page.
Limits
- Paywalled or offline sources are reported as "exists, claim not checked" rather than verified.
- Very old works, grey literature and personal communications may only be checkable by existence, not by content.
- Non-Latin scripts and transliterated names reduce match confidence; the report says when it is matching on partial metadata.
Questions
- Does citation verification work in Spanish and other languages?
- Yes. References in Spanish, English and the other major European languages are resolved against the same sources; matching handles accents, transliteration and mixed-language reference lists.
- Which citation styles are supported?
- APA, MLA, Chicago, Harvard, Vancouver, IEEE and inconsistent mixtures. ProofVerity parses the components of a reference, not its formatting.
- What if a reference is real but the page number is wrong?
- It is reported as misquoted, with the location where the claim was actually found when ProofVerity can confirm it.
- Can a student see why a citation was flagged?
- Yes. The student-facing evidence view shows the same status, sources checked and nearest match that the teacher sees.