Do AI Models Invent References? A Verification Audit of Citations in AI-Generated Academic Text

Working paper and evidence data. Five current model families (DeepSeek, Mistral, OpenAI, Anthropic, Gemini; 2026) audited under one protocol across 30 academic topics. Condition A verified all 194 author-year citations the models embedded in prose (none fabricated after adjudication). Condition B verified 1,500 one-shot full references against Crossref and OpenAlex with DOI resolution checks; 15.2 percent were fabricated or misattributed (family range 9.0 to 29.7 percent) and one in four supplied DOIs was defective. Files: per-reference and per-citation verdict evidence (JSONL/CSV), generation, verification, adjudication, and analysis code, figures, and the paper PDF. Full generated texts are not part of this release. Paper page: https://textpulse.ai/research/ai-hallucinated-references