A citation is supposed to be a promise: this claim rests on that source, and you can go check it. Large language models break that promise constantly. They generate references that look completely real, correct author names, plausible journal titles, formatted DOIs, for papers that were never written. This is called citation hallucination, and it is quietly becoming one of the biggest integrity problems in academic writing.

What is a hallucinated reference?

When an LLM is asked to support a claim with a citation, it does not look the source up. It predicts what a citation would probably look like based on patterns in its training data. Most of the time this produces something structurally perfect, real author, real-sounding title, real journal, and completely fictional. The paper simply does not exist, or it exists but says something different from what is being cited. Because the fabrication is fluent and well-formatted, it slips past writers, reviewers, and even editors who don't manually verify every reference.

How common is it, really?

The numbers, from peer-reviewed and journalistic sources, are worse than most researchers assume:

A widely cited study published in Scientific Reports found that 55% of citations generated by GPT-3.5, and 18% of citations from GPT-4, were entirely fabricated, with substantive errors in a large share of the "real" ones too (Walters & Wilder, Nature).
A 2026 analysis covering biomedical literature found that fabricated references have risen roughly 12-fold in two years: about 1 in 277 papers published in early 2026 cited a source that doesn't exist, up from 1 in 458 in 2025 and just 1 in 2,828 in 2023, with the sharpest jump coinciding with the mainstream adoption of AI writing tools in mid-2024 (STAT News).
A Deakin University study of mental health literature reviews found GPT-4o fabricated roughly 1 in 5 citations outright, with 56% of all citations either fake or materially wrong, and fabrication rates spiking above 30% for less-studied subtopics (StudyFinds).
A separate psychology-literature study found 32.3% of 300 AI-generated citations were hallucinated, concentrated heavily in niche subfields the model had seen little training data for (PsyPost).

These aren't hypothetical failure modes. Retraction Watch has documented journals actively retracting papers over fabricated, AI-generated citations, including a wave of retractions from Neurosurgical Review and a Springer Nature book pulled specifically for fake references (Retraction Watch). In at least one documented case, a paper rejected by one journal for fake references was resubmitted to another journal almost unchanged, and published anyway.

Why this is worse than a typo

A wrong page number is a correction. A hallucinated citation is a broken chain of evidence: a reader who follows it hits a dead end, a peer reviewer who trusts it approves a claim that was never actually supported, and a future paper that cites that paper inherits the fabrication. Because AI models produce hallucinations with the same confident tone as accurate output, there is no stylistic tell. The only way to catch it is to actually check every reference against a real source, which is exactly the step time pressure encourages people to skip.

The subfields most at risk are the ones with less training data, meaning niche, emerging, and non-English-language research is disproportionately vulnerable to being "cited" out of thin air.

How TeXposit prevents hallucinated citations

TeXposit is built around one principle: AI should assist your writing without being allowed to quietly invent your sources. Several features exist specifically to close the gap that causes citation hallucination:

Real reference search, not generated references. The built-in Reference panel searches live academic databases by title, author, or keyword and shows real citation counts and abstracts before you insert anything. Adding a citation pulls the actual BibTeX entry for that paper, it is never generated from an AI model's memory.
DOI import. You can pull a verified bibliographic record directly from a DOI, guaranteeing the metadata matches a real, indexed publication.
AI citation validity checking. TeXposit's writing assistant can check whether a citation it or you have added actually resolves to a real, existing source, and can fetch the underlying reference PDF to confirm the paper actually supports the claim being made, before you submit, not after a reviewer catches it.
Real-time citation verification. As you write, TeXposit continuously checks every \cite{} key against your .bib file and flags missing or mismatched references immediately, so broken citation chains never make it to a final draft.
Provenance certification. Every project carries a cryptographically signed, deterministic record of who wrote what, distinct from AI-hallucination risk, but part of the same underlying goal: making academic writing verifiable rather than merely plausible-looking.

The bottom line

Fabricated citations are no longer a rare edge case, they are showing up in published, peer-reviewed research at a fast-growing rate, and they are hard to catch by eye. Any workflow that lets an AI model draft prose without independently verifying its citations is exposed to this risk. TeXposit's citation tools are built so that every reference in your document traces back to something that actually exists, so you get the speed of AI-assisted writing without inheriting its biggest blind spot.