Why AI Invents Citations (and How to Catch Them)

If you've asked an AI chatbot for sources, you've probably been handed a citation that looks flawless but points to a paper that doesn't exist. It's not a bug you did something to trigger — it's a direct consequence of how these models work. Understanding why tells you exactly when to be suspicious.

Language models predict, they don't retrieve

A large language model generates text by predicting the most plausible next words, one after another, from patterns in its training data. It isn't looking anything up in a database. So when you ask for a citation, it produces the shape of a citation — a believable author, a real-sounding journal, a plausible year — without any record that the specific paper exists.

Why citations hallucinate so often

The model has seen millions of citations, so it's very good at the format — which is exactly what makes the fakes convincing. It confidently recombines real authors, real journals, and invented titles into references that are individually plausible but collectively fictional. This is called ungrounded generation: nothing ties the output to a real source.

When AI is most likely to make one up

• Obscure or very specific topics, where little real literature exists to draw on.
• Recent work published after the model's training cutoff.
• When you ask for “5 sources” — it will fabricate to hit the number rather than admit it's short.
• When you push for a citation to support a claim the literature doesn't actually make.

Grounded vs. ungrounded AI

Not all AI is equally risky. Tools that search a real database and cite only what they find — or that verify each reference against one — are far safer than a raw chatbot. The rule of thumb: if the tool can't show you where a source came from, treat every citation as unverified.

How to catch fabricated citations

The telltale signs (a DOI that won't resolve, a title that returns nothing on Google Scholar) are covered in How to Spot a Fake AI Citation. The fastest option is to check the whole list at once: Corroba Verify matches every reference against real scholarly databases and flags anything it can't confirm — without ever inventing a source to fill the gap.

FAQ

Do the newest models still do this? Less than older ones, but yes — any ungrounded model can fabricate. Always verify.

Why does it sound so confident? The model has no sense of whether a source is real; fluency isn't accuracy.

Can I just tell it to use only real sources? That reduces but doesn't eliminate fabrication. Verification is the only reliable fix.