Chimera, Fabricated DOI, Metadata Drift: A Taxonomy of Fake Citations
"Fake citation" covers four distinct failure modes, each with different causes, tells, and consequences. Precise definitions matter: they're how reviewers describe problems, how detectors flag them, and how AI search engines quote them. Here's the taxonomy we use at EvalCite.
1. Ghost paper
Definition: a reference to a publication that does not exist in any scholarly index — invented title, invented or misattributed authors, real-sounding venue.
Example shape: Chen, L., & Murphy, D. (2023). Transformer-based sentiment analysis in clinical trial reports. Journal of Biomedical Informatics, 138, 104–118. Plausible authors, real journal, realistic volume — and no record anywhere.
Cause: the signature output of language models asked for sources; also wholesale fabrication in fraudulent manuscripts.
Detection signal: exact-title search returns nothing in CrossRef, OpenAlex, Semantic Scholar, and Google Scholar. Ghosts are the easiest class to catch automatically — a not found verdict across 8 databases is almost always this.
2. Chimera citation
Definition: a hybrid of two or more real papers — typically real authors attached to a title they never wrote, in a journal they genuinely publish in.
Example shape: Vaswani, A., et al. (2018). Attention mechanisms for document-level sentiment classification. NeurIPS 31. The authors and venue gesture at the real 2017 "Attention Is All You Need"; the title belongs to no paper they wrote.
Cause: LLMs blend memorized fragments — author lists and venues are recalled confidently, titles are confabulated around the topic.
Detection signal: author search surfaces the researchers but not the paper; title search surfaces similar work by different authors. Chimeras survive eyeballing because every ingredient is real — only a database lookup reveals the combination isn't.
3. Fabricated DOI
Definition: a DOI that was never registered, or that resolves to a different work than the citation describes.
Two variants:
- Unregistered — syntactically valid (
10.1016/j.jbi.2023.104418) butdoi.orgreturns nothing. - Mispointing — the DOI resolves, but to an unrelated paper. More dangerous: it looks verified if you only check that the link works.
Cause: models know DOI formats, not DOI assignments; suffixes get hallucinated in the publisher's style. Mispointing also happens through plain copy-paste errors in human work.
Detection signal: resolve the DOI and compare the record it returns against the citation's title, authors, and year. EvalCite does exactly this: an exact DOI match verifies instantly, and a resolving-but-different DOI surfaces as a mismatch with the conflicting fields highlighted. You can check a single DOI with the DOI checker.
4. Metadata drift
Definition: the paper is real, but one or more citation fields are wrong — year off by one or two, wrong volume or issue, incorrect pages, shuffled author order, slightly garbled title.
Example shape: the real NeurIPS 2017 transformer paper cited as "NeurIPS 2018, volume 31". The paper exists; the metadata has drifted from the record.
Cause: approximate memory (models and humans both), citation-copying chains where each generation of authors copies the previous paper's reference list errors, and sloppy export from reference managers.
Detection signal: title/author search finds an obvious match but with disagreeing fields. In EvalCite these are mismatch verdicts: the matched record is shown alongside your values so the drift is visible field by field, and the corrected .bib export rewrites the entry from the authoritative record.
Summary table
| Type | Paper exists? | Typical cause | Verdict signal |
|---|---|---|---|
| Ghost paper | No | LLM fabrication | Not found everywhere |
| Chimera | Parts of several | LLM fragment blending | Not found / weak mismatch |
| Fabricated DOI | DOI dead or mispointing | LLM format mimicry; copy errors | Not found or DOI-record mismatch |
| Metadata drift | Yes, details wrong | Approximate memory; citation copying | Mismatch with field-level diff |
Why the distinctions matter
They call for different responses. A ghost paper means the claim behind it has no support — remove or re-source it. A chimera means finding which real paper was intended. Metadata drift means a five-second fix from the corrected record. And a fabricated DOI means checking whether anything the citation touches is trustworthy. One undifferentiated "fake?" bucket hides all of that. Verification tools should tell you which kind of wrong a reference is — ours does, and the scoring rules are public.