How to Check Whether a Source Really Exists (DOI, Crossref)

A reference exists if you can land on it from an identifier. It does not exist because it looks plausible, because the formatting is flawless, or because a chatbot repeated it twice. The check takes under a minute per source: resolve the DOI, and when there is no DOI, search a bibliographic database by first author and exact title. Everything else (a Google hit, a confident answer from a model, a tidy APA entry) tells you nothing about whether the work was ever published.

Below is the three-step routine, the failure mode most students miss, and what the research actually reports about how often AI-generated references are fake.

Step 1: Resolve the DOI before you read the reference

Take the DOI string from your reference and open https://doi.org/ followed by it. Three things can happen.

1. It lands on a publisher page whose title, authors and year match your reference. The work exists and your metadata is right.
2. You get "DOI Not Found". The identifier was invented or mistyped. Retype it once by hand, then discard the reference.
3. It resolves, but to a different paper. This is the one people miss. The link opens, the browser shows a real journal, and the student ticks it off. In fact the model borrowed a genuine DOI and wrapped an invented title and author list around it.

Outcome 3 is why "the link works" is not a verification step. Compare the resolved page field by field: surname of the first author, year, journal title, volume, page range.

Harvey and colleagues describe DOI use for uniquely identifying journal articles as essentially universal among publishers of scientific journals, supported by the registration agency Crossref, whose infrastructure was already more than a decade old when they wrote (s. 2). The practical consequence for your bibliography: for a journal article from roughly the last twenty years, having no findable DOI is a warning sign rather than a quirk.

Before: Thompson, R. (2019). Cognitive load in multimodal learning environments. Journal of Educational Psychology, 111(4), 512–528. https://doi.org/10.xxxx/edu0000341

After resolving: the DOI opens a genuine article in that journal, but by different authors and on a different topic. The reference goes in the bin, and so does the sentence that leaned on it.

Step 2: When there is no DOI

Books, book chapters, conference papers, most pre-2000 articles and a large share of humanities scholarship never received an identifier. Absence of a DOI is not evidence of fabrication there, so the check shifts to catalogues:

• Journal articles and chapters: search Crossref (search.crossref.org) or OpenAlex by first author surname plus three or four distinctive words from the title.
• Books: your national library catalogue, WorldCat, or an ISBN lookup. If the ISBN returns nothing, treat the entry the way you would treat a dead DOI.
• Anything paywalled: your university discovery service, which searches the databases your library subscribes to.

Search on the surname and a fragment, not the full title string. Kamińska explains why: reliable work on citation data depends on identifying a publication correctly, and the identifying key of a bibliographic unit is a composite one whose attributes vary with publication form, so a journal article, a monograph chapter and a whole monograph are matched on different fields (s. 97). She adds that databases carry both errors and duplicated records, and that even sophisticated analysis performed on erroneous, incomplete or duplicated data usually yields results of low reliability (s. 97–98). A single wrong word in a title can therefore sink an exact-string search on a source that genuinely exists.

One more caveat before you declare something fake. Birukou and Skalaban review how the same paper published in more than one language is handled, and find that different publishers may assign different DOIs to the different language versions of one article (s. 29–30). Their paper is about linking those identifiers rather than about student bibliographies, but the lesson carries: if your reference is a translation, a title that fails to match the resolved page may point to the original version rather than to a fabrication.

Step 3: Read the reference for the usual tells

The numbers here are worth knowing before you trust a model's reference list. Walters and Wilder had GPT-3.5 and GPT-4 produce short literature reviews on 42 topics and examined all 636 references in the resulting 84 papers. They found 55% of the GPT-3.5 citations fabricated against 18% of the GPT-4 ones, and among the references that pointed to real works, 43% and 24% respectively contained substantive errors (s. 1). Pooling six studies, they report that 51% of 732 citations were fabricated (s. 2).

Safran and Çalı ran a tighter test: four prompts on musculoskeletal rehabilitation, ten references with DOIs each. Of those 40 references, 7.5% were fully accurate and 42.5% were completely fabricated, with the remaining half partially correct (s. 695). They also note that fabricated references combine real author names, real journal titles and plausible publication years, which makes them hard to detect without thorough verification (s. 696).

So the tells are rarely stylistic. Look instead at:

• Numbers. Volume, issue, page range and year are where errors cluster.
• Real people, wrong paper. Recognisable names in a pairing that never co-authored anything.
• A real journal, an absent article. The venue checks out and the title returns nothing.
• Round, tidy page ranges on articles whose journal uses continuous pagination.

And one non-tell: asking the model whether its own citation is correct. Walters and Wilder point out that ChatGPT often gives incorrect responses to exactly that question (s. 2). Safran and Çalı asked the model to verify and revise its references, and the share of fully accurate ones rose from 7.5% to 77.5% (s. 695), which sounds like a fix until you notice that roughly one reference in five was still wrong afterwards. Self-verification improves a draft; it does not clear a bibliography.

Checking a whole reference list at once

Twenty-five references at a minute each is a manageable evening. The problem is that the check has to be repeated after every revision, and the failure mode is silent.

Szeider sets out the architectural principle that makes automated checking trustworthy: keep the language model out of the export path and fetch entries directly from the authoritative database, so the model can search and disambiguate but never retypes the metadata. Evaluated on 104 obfuscated citations, that design reached an 82.7% perfect match rate against 28.2% for standard web search, with zero cases of metadata corruption (s. 3). He also cites fabrication rates above 80% for certain publication categories in other models (s. 2).

That is the same rule behind cytado's bibliography checker: every entry is resolved against Crossref, OpenAlex and library records, and what comes back is the registered metadata rather than a paraphrase of it. If a source turns out not to exist, find-source looks for a real publication that supports the sentence you actually wrote. The academic references in this article were pulled from cytado's own corpus, with the page numbers read off the scanned pages.

The one-minute routine

1. DOI present? Resolve it, then compare author, year, title and pages against the landing page.
2. No DOI? Search Crossref or OpenAlex by surname plus title fragment; for books, use the catalogue or ISBN.
3. Nothing found? Check for a translated or renamed version before concluding it is fabricated.
4. Never ask the model that produced the reference to confirm the reference.

Related reading: Does AI invent sources? and how to find a source that supports your paragraph.