Researchers uncovered 295 conference papers with fabricated citations. Scanning over 17,000 submissions to the ACL, NAACL and EMNLP conferences from 2024‑2025, three Nara Institute scientists identified 295 articles containing at least one false reference, rising from 20 in 2024 to 275 in 2025—under 2 % of the total examined [4].
Advanced AI writing tools are a primary cause of the false references. New “AI scientist” systems introduced in late 2024 can search literature, suggest specific citations and draft review sections, accelerating manuscript preparation but also inserting non‑existent sources that later appear in published papers [1].
Erroneous database entries enable rapid citation spread. The study found “ghost entries” in major research databases; once an incorrect reference entered the system, it was copied into hundreds of subsequent articles, amplifying the problem [1].
Peer‑review overload hampers detection of bogus citations. Explosive growth in AI research, especially in fast‑moving subfields, leaves reviewers with limited time and expertise, forcing superficial checks that miss many fabricated references [1].
Co‑author Yusuke Sakai says reviewers cannot verify every citation under tight deadlines. He recounts completing ten reviews in one week, noting the process becomes “more formalistic than substantive,” and admits that even when he flagged a false reference it was not corrected [1].
Researchers propose automated screening and a continuous‑yearly review model. They suggest flagging papers with three or more suspicious references for manual verification and replacing the current conference‑centric evaluation with an ongoing review system akin to “megatidsskrifter” to improve reliability [1].