Why AI Invents Citations — and How to Catch It Every Time

Of all the ways AI can embarrass you academically, this is the one that does real damage: you cite a paper, your marker looks for it, and it does not exist.

It is worth understanding why this happens, because the explanation tells you exactly where to be suspicious.

What a language model is actually doing

A language model generates text that is plausible given everything before it. That is the whole job. It is not looking anything up, and it has no internal list of real papers to check against.

So when you ask for a citation, it produces something shaped like a citation. It has seen enormous numbers of real references, so it knows what they look like: plausible author surnames, a title using the right vocabulary for the field, a real journal, a sensible year, a well-formed DOI.

Every component is realistic. The combination is invented. This is why fabricated references are so much more convincing than fabricated prose — the format does the persuading for it.

Why “are you sure?” doesn’t help

Asking the model to confirm its own citation is close to worthless. It has no more access to the truth on the second attempt than the first, and it will often produce a confident reassurance — or, just as unhelpfully, apologise and invent a different fake.

The verification has to happen outside the conversation. There is no way around this.

The check

It takes about two minutes per source.

• Search the exact title in your library catalogue or a scholarly database. Use quotation marks. If nothing comes back, that is your answer.
• Check the DOI resolves. Paste it after https://doi.org/. A real DOI lands on a real paper. A fabricated one produces an error.
• Check the authors work in that field. Invented references sometimes borrow real, well-known names and attach them to work they never did.
• Open it. This is the one people skip. A real paper can exist and still not say what the AI claimed it says — and misrepresenting a real source is its own problem.

That last point deserves emphasis. The rule that survives every change in the technology is simple: never cite something you have not read. It was good advice before any of this existed and it remains the whole defence.

Where to be most suspicious

Fabrication gets more likely as the request gets more specific and more obscure. A model asked for “a foundational paper on natural selection” is on safe ground. A model asked for “a 2019 study on the effect of sleep deprivation on second-year engineering students in Scotland” is being asked for something that may not exist, and it will supply one anyway.

The more precisely your request describes a paper you wish existed, the more likely you are to be handed a fiction.

A better way to use AI in a literature search

Do not ask it for references. Ask it for search terms.

Describe your topic and ask which keywords, authors, debates and subfields you should be searching for. That plays to what the model is actually good at — knowing how a field talks about itself — while leaving the finding of real documents to a database that only contains real documents.

You can also work the other way round: find real papers first, then use AI to help you understand the ones you have already downloaded. A tool that answers from sources you supply cannot invent a reference, because it is only ever looking at what you gave it.