IF YOU’RE WONDERING
ChatGPT made up a study — are its citations real?
If an AI gave you a citation, a case, a statistic, or a study that turned out not to exist — or you’re about to rely on one and want to be sure — fabricated sources are one of the most common and best-documented failures these systems have, and the fix is a habit, not a better prompt. The unnerving part is that the fake ones look exactly like the real ones. We’re a research organization; this is squarely what we study, and the check below takes seconds.
The short answer: assume nothing, verify everything
A chatbot doesn’t look up a source and quote it. It predicts what a plausible citation looks like — a real-sounding author, a real-sounding journal, a real-sounding year — and assembles one. Most of the time the pieces are individually real and the combination is invented: a genuine researcher, a genuine journal, a paper that was never written. So a citation from an AI is a lead to check, never a source to cite. If you can’t find it outside the chat, assume it isn’t real.
Why it invents them so convincingly
The model is fluent in the shape of scholarship — the format of a DOI, the cadence of a case name, the house style of a journal. Producing that shape is exactly what it’s good at, whether or not a real document sits behind it. That’s why a fabricated reference carries the same confident polish as a real one, and why “it looked completely legitimate” is the single most common thing people say afterward. Confidence and formatting tell you nothing about existence. The underlying mechanism — why AI generates false material and presents it as fact — is on why AI makes things up.
This isn’t a fringe risk. Lawyers have been formally sanctioned by courts for filing briefs built on cases ChatGPT invented — the citations were perfectly formatted and simply did not exist. Students have turned in bibliographies of papers that were never written. The tool is genuinely useful for thinking and drafting; it is not a reference database, and treating it as one is where people get burned.
How to verify any citation in under a minute
- Search for it directly. Put the exact title (in quotes) into Google Scholar, PubMed, the journal’s own site, or a library database. If a paper, case, or study is real, it turns up. If it only exists inside your chat, it isn’t real.
- Check the DOI or link actually resolves. A made-up DOI 404s or lands on a different paper. Click it — don’t trust that it exists because it’s formatted like one.
- Ask the model to sort its own sources. Paste in: “For each source you just cited, tell me whether you retrieved it or generated it, and give me a link I can open. Flag any you cannot verify.” It will often withdraw the fabricated ones on the spot.
- Never let an AI citation reach anything that matters unverified — a filing, a paper, a medical decision, a published article. One search per source is cheaper than the alternative.
The broader five-minute conversation check is on Check your AI.
In one line: AI citations are leads, not sources. The fake ones look identical to the real ones, so the only reliable move is to search for each one yourself before you rely on it. If it doesn’t exist outside the chat, it doesn’t exist.
Where to go from here
Why it makes things up
What “hallucination” really is, why it isn’t lying, and why confidence tells you nothing.
The mechanism →Can I trust what it tells me?
The bigger question behind the fake citation: what a chatbot’s output is actually worth, and how to trust it wisely.
Did it lie to me? →Check the conversation
Six copy-and-paste prompts that make the AI account for its claims and sources. Five minutes.
Check your AI →Using AI for schoolwork
How to use it to learn without turning in sources that don’t exist.
Do it right →If the invented details were part of a bigger, flattering picture the AI was building — inventing figures to prove a point about your work or about you — that has its own page: why AI tells you you’re special.