A Taiwanese research team released an analysis of erroneous uses of artificial intelligence in academic articles, documenting cases where citation information produced by generative AI entered papers without any checking step.

Three types dominate: references to papers that do not exist; citations to real papers in support of claims they do not make; and entries in which authors and publication years are scrambled.

The cause lies in how generative models work. Language models are trained to produce plausible-looking strings of text, which makes highly formatted items such as bibliographic entries especially prone to invented combinations.

What the researchers emphasised was verification rather than the limits of the tools. Checking citations has always been basic work for authors and reviewers, and should be done regardless of whether AI was used. The problem is how often that step is now skipped.

Journals have begun to respond. Policies requiring disclosure of AI use have spread, and more are introducing automated reference-matching tools into the review process.

The issue is not confined to academia. The same class of error has been reported in journalism, legal filings and administrative documents. The shared difficulty is that the more correct the formatting looks, the harder the error is to spot.

The team proposed as follow-up work an examination of how frequency differs by discipline, since citation conventions and review practices vary across fields.