Hallucinated References in AI Papers: How CiteAudit Is Exposing Fake Citations in 2026 Conferences

Hallucinated References in AI Papers: How CiteAudit Is Exposing Fake Citations in 2026 Conferences

Hallucinated references — fabricated citations generated by large language models — are slipping through peer review at top AI conferences like NeurIPS and ICML in 2026. These AI-generated citations appear legitimate, complete with plausible authors, titles, and journal names, yet point to non-existent publications. According to The Decoder, up to 15% of recently accepted papers contain at least one hallucinated reference, often used to falsely bolster claims or mimic prior work.

How AI Models Generate Fake Citations

Large language models like GPT and Claude are trained on massive academic corpora, learning citation patterns so precisely they can invent convincing but fictional papers. Unlike simple typos, these hallucinations replicate formatting styles, use real-sounding journal titles (e.g., "Journal of Advanced Machine Learning"), and even include fabricated DOIs. This sophistication makes manual detection nearly impossible during fast-paced peer review cycles.

How CiteAudit Detects Hallucinated Citations

CiteAudit, the first open-source tool built to audit AI research citations, cross-references every reference against authoritative databases including arXiv, IEEE Xplore, PubMed, and Google Scholar. If a citation lacks a verifiable DOI or doesn’t return a match in any trusted source, it’s flagged with 92% accuracy. The tool also checks for mismatched author names, non-existent conference proceedings, and duplicate citations across papers — reducing false positives to under 3%.

Case Studies: Hallucinations in NeurIPS and ICML 2025

In a pilot review of 217 papers from NeurIPS 2025, CiteAudit identified 34 papers with hallucinated references — including one that cited a non-existent 2023 paper from "Nature AI" with a fake DOI. Another ICML submission referenced a "Stanford Research Group" that doesn’t exist. These cases confirm that AI-generated citations are not anomalies but systemic threats to research integrity.

Why This Matters: The Crisis of Academic Fraud in AI

Hallucinated references aren’t just errors — they’re a form of academic fraud. Researchers waste months chasing ghost studies, meta-analyses become unreliable, and trust in AI literature erodes. While tools like CiteAudit offer a technical fix, systemic change is overdue: journals must mandate citation verification, conferences should penalize intentional fabrication, and peer reviewers need training on LLM-generated deception.

Open access to CiteAudit — hosted on GitHub under an MIT license — empowers independent researchers, preprint platforms, and journals to audit submissions before publication. As AI-generated content becomes standard in drafting, the fight to preserve scholarly authenticity hinges on transparent, scalable tools. Hallucinated references must not become the new normal.