Can researchers stop AI making up citations? (www.nature.com)

OpenAI’s latest GPT-5 models mark notable progress in reducing “hallucinations”—the tendency of AI to fabricate citations or falsely claim completion of tasks—a notorious issue in large language models (LLMs). While improvements have lowered hallucination rates in many domains to levels “acceptable to users,” challenges persist in specialized fields like law and mathematics, and the model still struggles with some basic fact-based tasks. This is significant as hallucination affects the reliability of AI in academic and technical applications, especially in literature reviews where accurate citations are crucial. Technically, hallucinations arise from the probabilistic nature of LLMs, which predict plausible responses based on learned associations rather than verifiable facts, often guessing rather than admitting uncertainty. GPT-5’s gains come from scaling up training data, refining browsing capabilities to access up-to-date information, and incentivizing honest answers during training. Independent benchmarks show GPT-5 performs comparably or slightly better than human experts in citation-supported answers when online, but error rates increase markedly offline. Despite persistent hallucinations, GPT-5 is more transparent, reducing false task claims from 47% in previous models to 17%, a step toward safer and more trustworthy AI. Overall, while hallucination remains an intrinsic limitation, continued efforts to balance performance and honesty are key to deploying dependable AI tools for research and other high-stakes fields.