A fabricated citation from an AI tool is rarely random. It tends to look exactly like a real paper — plausible authors, a plausible journal, a plausible year — because the model is predicting what a citation should look like, not retrieving one that exists. Knowing the patterns makes them far faster to catch.
Why it happens
Language models trained to generate fluent text will produce a citation-shaped string even when no specific source backs it, because “a citation belongs here” is a stronger learned pattern than “I don’t actually have one for this claim.” This is a structural property of how these models generate text, not a rare glitch — assume it can happen with any tool that generates citations without retrieving them from a live source.
The tell-tale signs
A too-perfect fit — a citation that supports your exact claim with unusual precision is worth double-checking rather than celebrating. An unfamiliar but plausible-sounding journal name. An author whose other work you can’t independently find. A DOI or page range that looks formatted correctly but doesn’t resolve. Any one of these alone isn’t proof of fabrication, but two together are reason enough to verify before you cite.
The verification habit that actually works
Before a citation goes into a draft, confirm it resolves to a real, findable record — not just that the formatting looks plausible. A DOI that resolves, a listing in a database you trust, or the paper itself opening are the three checks that actually rule out fabrication; a well-formatted reference string alone rules out nothing.
Why retrieval-grounded tools behave differently
Tools that generate citations by retrieving from an actual indexed source — rather than generating a citation-shaped string from a language model’s training data alone — cannot fabricate a source that isn’t in the index, though they can still misattribute or misread one that is. That distinction matters more than which specific tool you use: know whether what you’re using retrieves or merely generates.
Build verification into the workflow
Keep the paper record beside the claim while drafting: open the DOI or source page, inspect the abstract or full text, and only then keep the citation. Retrieval-grounded tools make this easier when the citation remains connected to the selected paper; in the research workspace, a citation opens its corresponding source record so that this check is part of reading rather than a final clean-up task.