AI Models Cite the Wrong Source About 30% of the Time

A research team led by Professor James Zou at Stanford University published a study analyzing the information retrieval and citation accuracy of state-of-the-art AI models.

In tests of major models including GPT-4, Claude, and Gemini, RAG (Retrieval-Augmented Generation) systems maintained a high answer accuracy of about 85%, but when citing the sources underlying their answers, they pointed to irrelevant documents or incorrect sources about 30% of the time.

This phenomenon reveals a technical mismatch between text generation ability and actual citation ability. It raises the risk of spreading incorrect information, especially in fields such as medicine and law where verifying accurate evidence is essential.

The research team emphasized that for safe AI deployment, new technical standards are urgently needed to verify citation accuracy, going beyond simply getting the right answer.