Skip to content
humanise.ai Forum

How do I check that references from ChatGPT actually exist?

Staff answerWritten by Harriet Pell · Referencing and sources2 repliesupdated

Short answer

Look every reference up yourself in a place the AI does not control: your library catalogue, Google Scholar or the DOI resolver. Then open the source and read the part you plan to cite. ChatGPT can produce references that look perfect and do not exist, so a reference you have not seen is not yet a reference.

Written by Harriet Pell · Forum editor, voice and editing ·
Citehumanise.ai Forum, "How do I check that references from ChatGPT actually exist?", https://humaniseai.ai/forum/read/how-do-i-check-that-references-from-chatgpt-actually-exist/, accessed 2026-11-20.
  1. Forum editor, voice and editing ·

    Why this happens

    A chatbot predicts plausible text. A reference is a very predictable shape: authors, year, title in sentence case, journal, volume, pages, DOI. The model can assemble something in exactly that shape without any underlying article. King's College London's student guidance says generative tools "are NOT databases of knowledge", and warns that outputs can seem plausible yet contain "sources that do not actually exist".

    The APA Style team tested this in 2023. In their post on how to cite ChatGPT, they asked ChatGPT for a list of sources to back up its claims and got five references. They could find four online. The fifth did not seem to be a real article: the DOI belonged to a different article, and they could find nothing matching the authors, date, title and source it gave. They suggest scrutinising primary sources should become a standard process. That post is now kept for archival purposes only, but the advice about checking has not aged.

    Note the trap in that example. The fake reference had a DOI, and the DOI was real. It just pointed somewhere else. A reference can pass a quick glance and still be wrong.

    A five step check for every reference

    1. Search the exact title in your library catalogue or discovery tool. Use the university library search, not a general web search. If your library holds it, you will see the journal, volume and a full text link. Searching in quotation marks narrows it down.
    2. Try Google Scholar. Paste the title in quotation marks. A genuine article usually appears with the same authors and journal. If you find only similar titles by different authors, treat the reference as unconfirmed.
    3. Resolve the DOI. Type the DOI after https://doi.org/ in your browser. Check that the page you land on has the same title, authors and year as the reference. A DOI that opens a different article, or nothing at all, is a red flag.
    4. Compare every detail. Authors, year, journal name, volume, issue and page range should all match the publisher's page. Models often get one detail wrong, such as the year or the second author, even when the article is real.
    5. Open it and read the relevant part. This is the step that matters most. A reference can exist and still not say what ChatGPT claimed. Find the passage, check the claim, and note the page number.

    If a reference fails at any step, drop it. Do not "fix" it by guessing the closest real article. Search for a source yourself instead, using your reading list, library subject guide or Google Scholar with keywords.

    What about references that look right?

    APA's advice is worth taking literally: if the sources are real, accurate and relevant, it may be better to read the original sources, then paraphrase or quote them, than to rely on the model's interpretation. In practice that means treating an AI reference list as a list of search leads. Some will be real and useful. Some will be real but irrelevant. Some will not exist.

    Be most careful with:

    • statistics and quotations attributed to a named author;
    • references with very specific page ranges;
    • anything cited to support a claim that sounds surprisingly neat;
    • legal cases, standards and government reports, where one wrong digit points to something different.

    Keep a record

    For each source you keep, note where you found it (catalogue, database, publisher page) and the date. If a marker ever questions a reference, you can show where it came from. If you used AI to find leads, say so in your declaration if your university asks for one. The wider rules on what you may use AI for are in am I allowed to use ChatGPT for university assignments.

    If you find a bad reference after submitting

    Do not wait to be asked. Check your university's academic integrity policy and, where it allows, tell your module leader promptly. An honest mistake you report yourself is treated very differently from one a marker finds. Policies vary, so the wording on your own university site is what counts.

    If you want to test your own process, ask an AI tool for five references on a topic you know well and run the five steps on them. It takes about twenty minutes and teaches you what the failure looks like.

    5
    ReplyLink

Plain text or simple Markdown. Links are nofollow. Your email is never shown.

  1. Bronwyn AshbyStaffDetection and policy editor ·

    One more shortcut for the library step: most universities let you search by DOI directly in the discovery tool, which is quicker than typing a long title. And if your library does not hold an article, ask for an inter-library loan rather than citing it from an abstract alone. An abstract is not the paper.

  2. Harriet PellStaffForum editor, voice and editing ·

    A question that comes up: can I ask ChatGPT whether its own references are real? You can, but its answer is another prediction, not a lookup. It may insist a fake reference is genuine. Use it as a prompt to check, never as the check.

humanise.ai Forum. Questions and answers on using AI writing tools well at university and work, voice and tone, and AI detector worries.