Well-formed citation, no referent
- Id
- well-formed-citation-no-referent
- Status
- Active
- Severity
- high
- Detection
- judge
- Evidence grade
- corroborated
- Languages
- en
- Added
- 2026-08-15
- Updated
- 2026-08-15
Currently signals low-effort writing.
What it is
The citation has correct form and no referent. Case name, reporter volume and pin cite in the right order, resolving to no case. A DOI with a valid prefix that returns nothing. An ISBN whose check digit fails arithmetic.
Why it reads as machine-made
Citation form is one of the most patterned strings in written English, so it is cheap to produce. The referent is a fact about the world that the form does not carry, so it is not produced with it. Wikipedia built the failure into a speedy-deletion criterion, G15, which names unresolvable DOIs and invalid ISBN checksums among the signs a page could only have come from a language model. Courts have been dealing with the same thing in filings since 2023; a public database tracks the cases, though its counts disagree between summaries and the site refused our fetchers, so this entry cites its existence and no number.
Specimens
A 2019 review in the Journal of Applied Retail Analytics (doi:10.9999/jara.2019.04.007) found that shelf placement outperformed price promotion in 61 percent of trials.
We could not find a study comparing shelf placement with price promotion, so here is ours instead. Two stores, six weeks, the same product at eye level in one and at knee level in the other. Eye level sold 22 percent more units. Two stores is not a trial and we are not calling it a finding.
The journal, the identifier and the percentage in the first version are invented for this specimen. Figures in this repair are invented for the specimen.
How it is detected
- Rubric
- Task: check whether each citation resolves to something that exists. Step 1. Extract every citation: author, title, venue, year, identifier, page or pin cite. Step 2. Run the cheap arithmetic checks first. ISBN-10 and ISBN-13 check digits are computable without a network. So are volume and page ranges that fall outside a journal's published run. Step 3. Resolve each identifier. A DOI that returns no record, a case citation absent from the reporter, a URL that never existed rather than one that has moved. Step 4. Search title plus author independently of the identifier. A real work with a mistyped identifier will surface here. A fabricated one will not. Escape hatches, which return UNKNOWN rather than FLAG: paywalls and login walls; databases the reviewer cannot reach; works in languages or scripts the reviewer does not index; pre-digital material; preprints later withdrawn; personal communications and interviews, which have no public referent by nature and should be marked as such in the text. Decision. FLAG when a citation fails Step 2 arithmetic, or fails both Step 3 and Step 4. Mark UNKNOWN when only Step 3 fails and Step 4 could not run. Otherwise PASS. Output: one line per citation with its verdict, plus the count of FLAG results. A dead link and a fabricated citation are different findings and must be reported differently.
Who writes this way legitimately
Link rot kills real citations every day. Journals move, publishers merge, DOIs are registered late or never, and a working paper cited in 2021 can be unreachable in 2026 with nothing wrong about it. Legal and academic citation styles also vary enough that a correct reference can look malformed to a reviewer trained on a different one, and transliterated author names defeat naive search. The distinction that matters is between a referent that once existed and one that never did: an archive copy, a library catalogue record or a citing paper settles it. Reviewers working outside their own field should record UNKNOWN and ask, because absence of a search result is not absence of a work.
Model attribution
Documented across assistants rather than in one family. The behaviour follows from generating a reference rather than retrieving one, which is a property of the setup rather than of a vendor.
Platform notes
- wikipedia
- Criterion G15 for speedy deletion lists citations with invalid identifiers among the signs a page could only plausibly have been generated by a language model. It is a deletion criterion for pages, not a verdict about an editor.
Status history
| Date | Status | Rationale |
|---|---|---|
| 2026-08-15 | Active | Wikipedia's speedy-deletion criterion G15 names invalid identifiers as a sign, and court coverage records filings with citations that resolve to nothing. The pattern has an arithmetic component, which is rare in this dataset and makes it unusually cheap to check. |
Sources
- 01Wikipedia: Criteria for speedy deletion, G15communityaccessed 2026-08-14
- 02Charlotin, AI hallucination cases databasecommunityaccessed 2026-08-14
- 03Bloomberg Law, AI faked cases become core issue for judgespressaccessed 2026-08-14
CC BY 4.0 / The AI Tells Index, feedsquad.com/ai-tells