How to Stop AI Chatbots From Making Up Facts When You Ask for Sources
4 min read
You ask a chatbot for sources and it hands you a tidy list: authors, a journal, a year, even a link. Then you search for the first one and it does not exist. This is not a rare glitch. It is one of the best documented weaknesses of AI chatbots, and the honest answer to "how do I stop it?" is that you cannot fully stop it, but you can make it much less likely to hurt you.
How often do AI chatbots invent sources?
A 2023 study in Scientific Reports asked ChatGPT to write literature reviews and checked the references. It found that 55% of the citations from GPT-3.5 were fabricated, against 18% for GPT-4, and that many of the real ones still contained errors. A 2025 comparison of eight chatbots found that six of them made up a substantial share of the references they were asked for, while two (Grok and DeepSeek) made up none in that test, though none was fully accurate. Fabrication also varies by type: one study found far more invented references for journal articles than for books.
Two details matter for your routine. Fake references often look right: the same line of research reported that fabricated citations frequently came with DOIs, and many of those DOIs led to real papers on unrelated topics. And asking the chatbot "is this citation real?" is not a reliable check, because it can answer wrongly about its own output.
How to ask for sources so chatbots invent fewer
You cannot remove the risk, but wording helps.
- Use a mode that searches. A chatbot that browses or searches live pages can link to pages that exist, which is safer than a model writing a reference from memory. It still needs checking.
- Ask for what the model can stand behind. "Name the main viewpoints on this topic and the search terms I should use" is safer than "give me five peer reviewed sources with DOIs."
- Paste the source in. If you give the chatbot the article or the passage, ask it to answer only from that text and to quote the line it used. It cannot invent a source you already supplied.
- Allow "I don't know." Add: "If you are not sure a source exists, say so instead of guessing."
- Ask for the claim, not just the citation. "Which sentence in the source supports this?" exposes a made-up reference quickly.
How to check an AI-given source in two minutes
Treat each AI-suggested source as a lead, not a fact.
- Search the exact title in Google Scholar, your library, or the publisher's site.
- Check the DOI by pasting it into doi.org or Crossref. Confirm it opens the same title, not a different paper.
- Match the details: authors, journal, year, volume and pages.
- Open the source and find the claim. A real paper can still say something different from what the chatbot claims.
- Drop what you cannot find. If a source does not exist, do not keep it "just in case."
A worked example of the two-minute check
Suppose you ask a chatbot for studies on sleep and memory and it lists a paper by a named author in a named journal with a DOI. Search the exact title: if nothing comes up, stop there. If it exists, paste the DOI into doi.org and see whether it opens the same title. Then read the abstract and ask whether the paper actually reports what the chatbot said. A paper on sleep in mice does not support a claim about students, however real the citation is. Numbers also need care: when the chatbot quotes a figure, find it in the paper. For a related case, see using ChatGPT for financial modeling without hallucinated numbers.
Use AI for searching, not for sourcing
The safest pattern is to let the chatbot help you find where to look (keywords, related topics, how experts phrase the question) and then do the sourcing in a real database. Use it to explain a paper you already have, to summarize a passage you pasted, or to turn your notes into an outline. Do not use it as the final authority for a reference list.
Sources
- Walters and Wilder, Fabrication and errors in the bibliographic citations generated by ChatGPT, Scientific Reports (2023)
- Assessing the performance of 8 AI chatbots in bibliographic reference retrieval, arXiv (2025)
- Reference Hallucination Score for medical AI chatbots, JMIR Medical Informatics (2024)