If an AI made up a source in your draft, treat every other citation in that draft as unchecked. The cause was the request, not one unlucky line. Open each source, confirm it says what the sentence claims, and either replace or delete the claims that fail. Then change how you ask: give ChatGPT, Claude, Gemini or Copilot the sources and ask for a quote behind each claim, instead of asking it to remember sources for you.
What to do in the next hour
- Stop the draft where it is. Fixing the one bad reference and sending the rest assumes the others were produced differently, and they were not.
- List every citation and every claim that leans on one. The Fact Check List does this for you: paste the text and it pulls out each source, number, quote and date, riskiest first, with a search link for each.
- Find each source yourself. Search for the title and author rather than clicking the link the model gave, because a model can attach a real link to the wrong paper.
- For every claim that fails, find real support or cut the claim. Do not swap in a source that is close enough. That is how an invented citation turns into a misquoted one.
- Note what you changed, so whoever reviews the draft can see the gap was found rather than papered over.
Why models invent sources
A chat model writes the most likely next words. A reference has a very predictable shape (authors, year, title, journal, pages) and the model has seen millions of them. Asked for sources it was not given, it produces text in that shape. Sometimes the text matches a real paper. Sometimes it joins a plausible author, a plausible title and a real journal into something that was never published. Nothing in that process checks the result against a library.
Our guide to reducing AI hallucinations covers the general mechanism. For citations the point is narrower: “give me sources” is a request a model can only answer from memory, and its memory of references is a memory of what references look like.
It invents most where you know least
The clearest recent measurement is a 2025 study in JMIR Mental Health. Researchers at Deakin University asked GPT-4o for six literature reviews and checked all 176 citations it produced. 35 were fabricated. Of the 141 real ones, 64 had at least one error, most often a wrong or broken DOI. Only 77, under half, were real and correct.
- Fabricated20%
- Real, with errors36%
- Real and correct44%
Source: Linardon et al., JMIR Mental Health, 2025.
The more useful result is where the inventions clustered. For major depressive disorder, a heavily studied topic, 6% of citations were fabricated. For binge eating disorder and body dysmorphic disorder, both less studied, the rates were 28% and 29%. A review of one specialized angle, digital treatments for binge eating disorder, reached 46%, against 17% for a general overview of the same condition.
The narrower and less studied the topic, the more the model invented. Source: Linardon et al., JMIR Mental Health, 2025.
That is the trap. People ask a model for sources when they do not know the literature themselves, which is exactly where it is most likely to invent, in a format nobody can tell apart from the real thing at a glance.
Three ways a citation fails
Checking a citation means asking three questions in order. Each failure has a different fix, and the third one is the failure people miss, because the source is real and the link works.
In the study above, 35 citations failed the first question and 64 failed the second. Nobody counted the third, which needs a person to read the source.
Give it the sources, ask for quotes
The fix is to stop asking the model to remember. Supply the sources, ask it to show the sentence behind every claim, and give it permission to say that a source does not support something. Here is the same request both ways:
Write a paragraph on how remote work affects productivity, with sources.
Write one paragraph for our leadership team on what the evidence says about remote work and productivity. Under 150 words. Use only the three sources pasted below, numbered 1 to 3. After each claim, put the source number and quote the sentence that supports it. If the sources do not support a point, say so instead of adding one, and never cite anything that is not below. A good paragraph lets a reader check every claim in under a minute. Sources: """ [1] ... [2] ... [3] ... """
The scores come from the free checks in the Prompt Grader, so you can paste both and get the same numbers. A score measures the brief, not the facts. What matters for citations are two lines: “use only the sources below” and “say so instead of adding one.” The first removes the need to remember. The second removes the pressure to fill a gap.
If you do not have sources yet, find them first, then write. The deep research request template briefs a research tool to bring back sources you can open, and summarize what several studies say works once you have them. To check a finished draft against your sources, use check a draft’s claims against your sources. For a long report, answers from a long document with page references gives you a page number to verify instead of a citation to trust.
To make this the default in every chat, add one line to your standing instructions: never cite a source I have not given you, or one you did not open in a search during this chat. The custom instructions generator writes that for ChatGPT, Claude, Gemini or Copilot and says where to paste it.
Does web search fix it?
The obvious objection is that this is yesterday’s problem. ChatGPT, Claude, Gemini and Copilot can all search the web now, so their sources are real pages. That is partly true, and it changes which failure you meet more than whether you meet one.
In March 2025 the Tow Center for Digital Journalism gave eight AI search tools excerpts from real news articles and asked each to name the article, publisher, date and link. Across 1,600 queries the tools were wrong more than 60% of the time, from 37% for Perplexity to 94% for Grok 3, and more than half of the responses from Gemini and Grok 3 cited fabricated or broken links (Columbia Journalism Review). Tools have improved since, but the shape of the error matters more than the rate: search swaps invented sources for real ones that may not say what the answer claims. That makes the third question in the tree the one to keep asking.
If it already went out
Correct it the way you would correct a wrong number: say which reference was wrong, what replaces it or that the claim is withdrawn, and whether any conclusion changes. Then check every other citation in the document before anyone else does. Courts are one place where this has become routine. Damien Charlotin’s AI Hallucination Cases database, which tracks court decisions that deal with AI-generated errors such as invented case law, listed 2,231 decisions on October 28, 2026. A short correction you send costs less than one someone sends you.
The checklist
- A person has opened every source in the draft, not just clicked the link.
- Each source exists, its details match, and it says what the sentence claims.
- Claims without support are cut, not moved to a source that sits nearby.
- The prompt supplied the sources, asked for a quote per claim and allowed “not supported.”
- Answers from search tools went through the same three questions.
- Anything already sent has a correction that names what changed.
Key takeaways
- One made-up source means every citation in the draft needs checking, because they all came from the same request.
- Models invent references by imitating their shape; nothing in the process checks that they exist.
- In a 2025 study, GPT-4o fabricated 35 of 176 citations, and fabrication rose from 6% on a well-studied topic to 46% on a narrow one.
- Check each citation in order: does it exist, do the details match, does it say what you claim.
- Supply the sources, ask for a quote behind every claim, and allow the model to say a point is not supported.
- Web search reduces invented sources but not misattributed ones, so keep reading the source.
Questions
Why does ChatGPT make up sources?
Because it writes text in the shape of a reference without checking that the reference exists. Asked for sources it was not given, ChatGPT, Claude, Gemini or Copilot produces plausible authors, titles and journals from memory. In a 2025 study of GPT-4o, 35 of 176 citations were fabricated. Give it the sources and ask for a quote behind each claim instead.
How can I tell if a citation from AI is real?
Search for the title and author yourself rather than clicking the link the model gave. Check that the year, journal and DOI match, then find the sentence in the source that supports the claim. The Fact Check List pulls every citation, number and quote out of an AI answer, with a search link for each.
Does AI with web search still make up citations?
Less often, but it still attaches answers to the wrong source. In a 2025 Tow Center test of eight AI search tools, the tools named the wrong article, publisher, date or link on more than 60% of 1,600 queries. Open every link and find the sentence that supports the claim before you rely on it.
What should I do if I already sent a report with a fake source?
Send a short correction: which reference was wrong, what replaces it or that the claim is withdrawn, and whether any conclusion changes. Then check every other citation in the report, because one invented source means the rest came from the same request. The guide to fact-checking AI answers sets out the order to check them in.