ACL-family conference papers show rising hallucinated citations

Summary

Sakai, Kamigaito, and Watanabe analyze ACL, NAACL, and EMNLP papers from 2024 and 2025 and report nearly 300 papers with at least one hallucinated citation. The preprint says the issue was concentrated in 2025, especially EMNLP 2025, including more than 100 main-conference and Findings papers.

AI role

Language models are the suspected source of plausible but non-existent references that entered accepted computational-linguistics papers.

Narrative role

This provides field-specific evidence that hallucinated references are not only a general publishing problem but have entered mainstream AI and NLP conference outputs.

Caveat

The work is labeled work in progress and is a preprint, so the exact counts and detection methodology may change with later versions or venue corrections.

Related events

• Large-scale audit finds hallucinated citations entering scientific literature
• Reference hallucination becomes a measurable LLM failure mode