G1Text analysisHallucinationLayer 1 (Deterministic)

Citation Hallucination

Detects citations that appear fabricated, references with plausible-sounding but non-existent authors, journals, or publication details that AI models tend to invent.

Technical description

Extracts all citation references from the text and analyzes them for hallucination markers: implausible author name combinations, non-existent journal names, impossible publication dates, inconsistent citation formatting within the same paper, and references that cite well-known authors with wrong paper titles. Uses heuristic checks including journal name validation against known databases and author name pattern analysis. Now also flags placeholder citation stubs ([CITATION], (Author, Year)) and fabricated or placeholder DOIs, and labels each finding by fabrication category (Placeholder Hallucination, Identifier, and so on) following the NeurIPS-100 five-category taxonomy. Verification against external databases is handed to L3.

How it works

Layer 1 (deterministic): Extracts citation entries using patterns for common citation formats. Checks author name patterns for statistical implausibility. Validates journal names against a known journal list. Checks for date impossibilities and format inconsistencies. Flags citations with multiple red flags.

Why this matters

Citation hallucination is one of the most serious forms of AI-generated academic fraud. Language models frequently invent plausible-sounding but completely fictional references, including fake authors, journals, and DOIs. These fabricated citations can mislead readers and undermine the integrity of the academic literature if not detected.

Score thresholds

0-1: All citations appear legitimate and well-formed
2-3: Some citations have minor irregularities
4-5: Multiple citations show strong hallucination markers

Limitations

Unusual but legitimate journals may be flagged as suspicious. Authors from underrepresented regions may have name patterns that differ from the training data. Preprints and working papers may not appear in journal databases.