AI Hallucinations: Models Cite Wrong Sources, New Study Shows

AI models often provide correct answers but cite the wrong sources. This issue, called "attribution hallucination," is revealed by a new benchmark named CiteVQA. For example, Gemini-3.1-Pro-Preview scored 76 out of 100 when precise citations were required. GPT-5.4's performance dropped from 87.1 for answer quality to just 59 when correct sourcing was needed. Open-source models scored even lower. This matters because in fields like finance or healthcare, every claim needs a verifiable paper trail, and incorrect sourcing could have serious consequences.