Hallucinated Citations
Bibliographic entries generated by large language models during AI-assisted academic writing that are syntactically well-formed but refer to non-existent sources, including fabricated author combinations, fictitious journal names, and nonexistent page numbers. This phenomenon represents an academic integrity challenge associated with AI-assisted writing.
Core Facts
Timeline (last 90 days)
When generating text, AI models may "fabricate" plausible-looking citations, including authors, titles, dates, and URLs, based on patterns in the training data; this phenomenon is called "hallucinated citation."
Hallucinated references are a common error of large models: when external resources are needed, the model generates URLs, library names or APIs that are correctly formatted and seemingly plausible but do not actually exist; in interactive web pages that need to load 3D models, textures or third-party scripts, fake links directly break the rendering pipeline, causing sliders, dates and values to lose their linkage.
In the 3D explanation question on the causes of the seasons, Gemini 4 Argon referenced a URL that does not exist, the 3D scene would not load, the slider could not be dragged, and the date and numbers stayed forever at June 21.
Last year, ICLR desk-rejected some submissions due to hallucinated citations.
Some conferences and journals have begun requiring authors to submit accessible links to cited sources, or having reviewers specifically check reference lists, in order to address the problem of hallucinated citations.
This phenomenon is referred to as hallucinated citations or parametric knowledge leakage.
Even when the model is explicitly required to answer only based on retrieved documents, it still fills in gaps using parametric memory and generates seemingly well-formatted citations.
Retrieval-Augmented Generation (RAG) reduces the probability of hallucinated content and enables traceable citations by anchoring model outputs to a verified document database.
All Facts (8)
Retrieval-Augmented Generation (RAG) reduces the probability of hallucinated content and enables traceable citations by anchoring model outputs to a verified document database.
65%UnverifiedWhen generating text, AI models may "fabricate" plausible-looking citations, including authors, titles, dates, and URLs, based on patterns in the training data; this phenomenon is called "hallucinated citation."
50%UnverifiedHallucinated references are a common error of large models: when external resources are needed, the model generates URLs, library names or APIs that are correctly formatted and seemingly plausible but do not actually exist; in interactive web pages that need to load 3D models, textures or third-party scripts, fake links directly break the rendering pipeline, causing sliders, dates and values to lose their linkage.
50%UnverifiedIn the 3D explanation question on the causes of the seasons, Gemini 4 Argon referenced a URL that does not exist, the 3D scene would not load, the slider could not be dragged, and the date and numbers stayed forever at June 21.
50%UnverifiedSome conferences and journals have begun requiring authors to submit accessible links to cited sources, or having reviewers specifically check reference lists, in order to address the problem of hallucinated citations.
50%UnverifiedThis phenomenon is referred to as hallucinated citations or parametric knowledge leakage.
50%UnverifiedEven when the model is explicitly required to answer only based on retrieved documents, it still fills in gaps using parametric memory and generates seemingly well-formatted citations.
50%UnverifiedLast year, ICLR desk-rejected some submissions due to hallucinated citations.
50%