Why does ChatGPT invent references?
Because a language model writes references the same way it writes everything else: by predicting what plausibly comes next, not by looking anything up. A citation that looksexactly right is a success by the model's own measure — whether or not the paper exists.
It predicts text, it doesn't retrieve facts
A model trained on millions of papers has absorbed the shapeof a citation: a handful of surnames, a noun-phrase title, a journal that fits the topic, a plausible year, a DOI with the right prefix. When you ask for sources, it assembles a reference in that shape. Nothing in that process checks a database — so the output is a confident average of “what a citation about this topic tends to look like,” which is often a paper that was never written.
Why the fakes look so convincing
The same training that lets a model mimic citation formatting also lets it borrow real author names, real journals, and real DOI patterns. The result is a reference that is individually plausible in every part — which is exactly what makes it hard to catch by eye. The authors might be real researchers in the field; the journal might exist; the DOI might even resolve — just to a different paper.
Doesn't web search fix it?
Partly. Tools that retrieve real sources before answering (retrieval-augmented or “AI search” systems) fabricate less, because some references are now grounded in a real lookup. But they still summarize, paraphrase, and stitch together — and a retrieved source can be attached to the wrong claim, or a fabricated reference can slip in alongside the real ones. Grounding lowers the rate; it doesn't drive it to zero.
How often does it happen?
Measured fabrication rates in model-generated bibliographies have ranged widely across studies, models, and prompts — from roughly one-fifth to most of the references in some experiments. The rate depends heavily on how obscure the topic is and how specific the prompt is: the less the model “knows,” the more it improvises. The practical takeaway isn't a single number — it's that you cannot assume a model-supplied reference is real.
What actually reduces the risk
Verify every reference against authoritative databases before you rely on it — confirm the work exists, that the identifiers point to it, and that the metadata matches. That's the check Hallucite automates across a whole bibliography at once. See how verification works or read what a fake citation is.