When AI Writes the Bibliography, Peer Review Isn't Enough
Agentic AIThe peer review process at top AI conferences is under stress — and not from too many submissions. It’s from fabricated citations that are passing review undetected.
GPTZero’s investigation identified more than 50 hallucinated references in ICLR 2026 submissions alone. The papers are real. The research is real. But the citations anchoring the related work sections point to sources that don’t exist.
The Verification Gap
The problem isn’t that researchers are being deliberately deceptive. The more likely explanation is systematic: LLM-assisted writing produces fluent, plausible-sounding citations because fluency and accuracy are different skills. A model trained on academic text learns to produce well-formatted references without any mechanism for verifying that the referenced work exists.
What’s notable is that the problem persists through peer review. Human reviewers, who are already overburdened, rarely verify bibliographies citation by citation. The assumption of good faith that underpins peer review creates a gap that hallucinated citations exploit by default.
The verification gap compounds in a specific way: as AI-assisted writing becomes standard, reviewers face an asymmetric problem. Authors who use LLMs to draft related work sections produce denser bibliographies faster, increasing the absolute number of citations a reviewer would need to verify — at exactly the moment when verification is least likely to happen.
CiteAudit: The Automated Counter
The new open-source response to this problem is CiteAudit, which uses five specialized agents in sequence: an extractor pulls bibliographic details from PDFs, a memory agent checks against previously verified citations, a web search agent queries Google, a judge agent compares citations against retrieved evidence, and a scholar agent cross-references authoritative databases including Google Scholar and ArXiv.
The system processes 10 references in 2.3 seconds and achieves 97.2% overall accuracy — substantially better than direct LLM verification, which misses roughly 22% of real hallucinations while falsely flagging 48% of legitimate citations. Speed and specificity are the differentiating factors. CiteAudit runs locally with no token costs and is available as a free web application at checkcitation.com with a 500-citation daily limit.
For context on how it was evaluated: the benchmark dataset contains 6,475 real citations and 2,967 fabricated ones, sourced from Google Scholar, OpenReview, ArXiv, and BioRxiv.
The So What
The citation integrity problem in AI research is a preview of what happens when AI-assisted writing scales into any domain that relies on cited evidence — legal briefs, medical literature reviews, financial due diligence. The mechanism is identical: LLMs produce plausible-sounding references, humans trust the formatting, and no one checks.
CiteAudit represents a tractable response to a specific and growing problem. Conferences should consider integrating automated citation verification into submission pipelines before acceptance — not as a gatekeeping mechanism, but as a quality signal that reduces reviewer burden. The tool exists. The question is whether the institutional will to use it follows.
For practitioners using AI to assist with research or writing, the immediate takeaway is straightforward: verify every citation before it leaves your hands. A hallucinated reference that passes peer review at a top conference is embarrassing. One that passes through a legal filing or clinical summary carries higher stakes.
Recommended Reading:
Photo by Tima Miroshnichenko on Pexels
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →