What chatbots get wrong
ChatGPT, Claude, Gemini, and similar tools compose a citation the way they compose a sentence: by continuing a pattern. The usual failures are a DOI that resolves to a different paper, a real author attached to a title they never wrote, and volume, issue, and page numbers that belong to another article.
Turning on browsing or a research mode does not finish the job. A model can still pair a real DOI with the wrong title. Treat every line the chat produced as unchecked until a bibliographic database agrees with it.
What RefCheck does with that list
Paste the references the chat gave you, one per line. RefCheck parses each one and searches Crossref and OpenAlex first. Biomedical lines also go to PubMed and Europe PMC, preprints to arXiv, computer-science lines to DBLP, datasets to DataCite, and books to Open Library and Google Books. Semantic Scholar is part of the same cross-check.
You get one of four verdicts: real, real with errors, likely fabricated, or unconfirmed. “Real with errors” means a matching work was found but the year, pages, venue, or DOI in your line disagrees with the record. The screen shows a corrected citation. It does not decide whether that paper supports the sentence in your draft.
What a miss does not prove
“Likely fabricated” means the open databases RefCheck queries did not return a matching work. “Unconfirmed” means the evidence was too thin to call, which is common for Chinese-journal articles, very new preprints, and some books. Neither verdict is a plagiarism score, and neither replaces reading the paper.
The same steps apply if the list came from DeepSeek, Kimi, or another assistant. The Chinese page for this job is ChatGPT 参考文献核查. For the general idea of a made-up citation, see the hallucinated citation checker.