Why Structural Checking Catches What Reading Misses
A reference fabricated by ChatGPT, Gemini, or Claude is constructed to pass a visual read. It has a plausible author surname, a journal title that sounds real, a year that fits the paper's timeline, and often a DOI-shaped string that looks correct at a glance. The problem is structural, not stylistic — the fabricated entry is missing the one thing that cannot be faked convincingly: a working, traceable identifier that resolves to an actual document. Reading a reference list catches typos and obvious nonsense. It does not catch a citation that is internally consistent but externally unverifiable.
This is why the check works field by field rather than holistically. Each reference is tested against five specific requirements — author, year, title, publisher or journal, and an identifier — and the result is a count, not an impression. A reference with four of five fields present but no DOI and no URL is flagged regardless of how convincing the rest reads, because the missing field is the one that actually matters for verification.
The Five-Field Requirement
Every reference entry is tested against the same five fields, regardless of citation style. The fields are not weighted equally — a missing identifier carries more risk weight than a missing field name, because it removes the only path to independent verification.
- Author — a recognizable name pattern at the start of the entry, in any of the formats used across APA, MLA, Chicago, Harvard, or Vancouver conventions.
- Year — a four-digit publication year, checked against a plausible range rather than assumed present.
- Title — either a quoted phrase or a sentence-length fragment long enough to be a real title rather than a placeholder.
- Publisher or journal — a named source, detected through journal-volume patterns, publisher-and-place patterns, or explicit keywords like Press or Journal of.
- Identifier — a DOI matching the valid 10.XXXX/ format, or a working URL. This is the field most often missing in fabricated entries.
Citation Style Variance Across Disciplines
Different academic fields default to different citation conventions, and a checker that only recognizes one style will misflag a correctly formatted paper from a different discipline. Psychology and education papers default to APA. Literature and humanities papers default to MLA. History papers typically use Chicago Notes-Bibliography. Medicine and the biomedical sciences use Vancouver, identifiable by its numbered in-text markers and Lastname Initial author format with no comma. Engineering and computer science papers typically use IEEE, also numeric but with a distinct reference-list format.
A checker that assumes APA formatting will misread a Vancouver-style author field — Patel R, Kumar S looks malformed under APA rules but is correctly formatted under Vancouver convention, where authors are listed as surname followed by initials with no comma and no period. Recognizing which style a paper uses before evaluating field completeness prevents this exact false positive.
What Gets Flagged, and Why It Does Not Mean Fabrication
A flagged reference means one or more of the five fields could not be verified as present — it does not mean the source is confirmed fake. A real book published before publishers routinely assigned DOIs will correctly show no identifier and will be flagged for that reason alone. The flag is a prompt to check manually, not a verdict. The distinction matters because the goal is catching the small number of entries that need a second look, not generating a long list of false alarms that gets ignored.