Why ChatGPT Makes Up Citations (and How to Get Real Ones)
You ask an AI for sources, it hands you a clean list of references — authors, journal, year, even a DOI. They look perfect. The problem: some of them were never published. Here's why it happens and how to write with citations you can actually verify.
The short answer
ChatGPT and similar chatbots don't "look up" papers. They generate text one word at a time by predicting what usually comes next. A citation is just a very predictable pattern — Surname, Initials (Year). Title. Journal, Volume(Issue), Pages. The model can reproduce that shape flawlessly while filling it with a paper that does not exist. This is called a hallucinated citation, and it is one of the most common — and most dangerous — failure modes of AI writing.
A citation that looks real and a citation that is real are two completely different things. A language model is optimized for the first, not the second.
What a hallucinated citation actually looks like
Fabricated references are convincing precisely because the AI has read millions of genuine ones. A typical fake includes:
- Real, well-known authors — often researchers who publish in that field, attached to a paper they never wrote.
- A real journal — the correct venue for the topic, with a plausible volume and issue.
- A syntactically valid DOI — the right format (
10.xxxx/…) that nonetheless resolves to nothing. - A title that fits the argument too neatly — because it was invented to support the sentence, not the other way around.
This isn't hypothetical. In a widely reported 2023 U.S. court case, two lawyers were sanctioned after submitting a legal brief containing several judicial decisions that ChatGPT had invented wholesale — complete with fake quotes and citations. If trained professionals were fooled, a student skimming a reference list certainly can be.
Why language models invent references
To understand the fix, it helps to understand the cause. Three things combine:
1. Next-token prediction, not retrieval
A base language model has no live connection to a database of publications. It has patterns absorbed during training. When you request a source, it generates the most statistically likely citation-shaped text — regardless of whether that specific paper is real.
2. Training data doesn't store papers as facts
The model may have seen a paper's abstract during training, but it doesn't keep a verifiable index of "this exact title by these exact authors exists." Details blur together, so it confidently mixes a real author with a plausible-but-fake title.
3. It is rewarded for sounding helpful
Chat models are tuned to be fluent and confident. "I'm not sure that source exists" is rarely the response you get. Fluency and confidence are exactly what make a fabricated citation so easy to trust.
How to tell if a citation is fake (a 60-second checklist)
Before you paste any AI-suggested reference into your paper, run these checks:
- Resolve the DOI. Put it after
https://doi.org/. A real DOI opens the paper's page; a fake one 404s. - Search the exact title in Google Scholar or your library. Zero results = red flag.
- Match the authors to the journal. Do these authors actually publish there? Does the volume/year line up with the journal's real history?
- Open it and read the abstract. If you can't find the actual paper, you can't cite it — full stop.
If a reference fails any of these, don't "fix" it — remove it. A missing citation is a small problem; a fabricated one is misconduct.
Never cite a source you have not opened. If the AI can't show you the real paper behind a claim, treat that claim as unsupported.
How to write with citations you can actually verify
The good news: hallucinated citations are a solvable problem. The fix is to change how the AI is allowed to cite.
Use grounded (retrieval-based) AI, not a raw chatbot
A grounded tool first retrieves a set of real papers from a live index, then writes using only those papers — and can only cite sources that were actually pulled in. Because the model operates in a "closed universe" of verified sources, it has nothing to invent. This is exactly how CiteGood works: every sentence is generated from real literature, and every citation links back to a source you can open.
Verify claim-by-claim, not just the reference list
Even with real sources, make sure each source actually supports the sentence it's attached to. A real paper cited for a claim it never made is still wrong.
Format last, with a citation generator
Once your sources are real and verified, formatting is the easy part. Use our free APA / MLA / Chicago / IEEE citation generator to turn a DOI or paper details into a correctly formatted reference in seconds.
Write grounded. Cite real sources.
CiteGood is the AI academic writer that only cites literature that actually exists — no hallucinated references, every claim linked to a source you can verify.
Try CiteGood free →Frequently asked questions
Why does ChatGPT make up citations?
Large language models generate text by predicting the most likely next word, not by looking anything up in a database of papers. When you ask for sources, the model produces text that statistically resembles a real citation — plausible authors, a believable journal, a normal-looking DOI — without any guarantee that the reference actually exists.
How do I know if an AI-generated citation is real?
Check the DOI (paste it after https://doi.org/ — it should resolve to the real paper), search the exact title in Google Scholar, and confirm the authors and journal match. If the DOI 404s or the title returns no results, the citation is almost certainly fabricated.
Do all AI writing tools fake citations?
No. Tools that generate freely from the model's memory (like a raw chatbot) are prone to it. Tools that use retrieval or 'grounding' — pulling from a fixed set of real papers and citing only those — do not invent references, because the model is only allowed to cite sources that were actually retrieved.
Is it plagiarism to use AI citations?
Fabricating references is a form of academic misconduct in its own right, separate from plagiarism — you are attributing claims to sources that don't exist. Using AI to help you write is generally allowed if your institution permits it, but every citation must point to a real, verifiable source.