Despite rapid progress in developing large language models (LLMs), artificial intelligence (AI) hallucinations—instances where AI generates plausible but factually incorrect responses—continue to pose significant challenges. Professor Yun-Nung Chen from National Taiwan University's Department of Computer Science highlighted the need for increased user participation and greater efforts by AI systems to express uncertainty and cite sources to enhance trustworthiness.
AI hallucination occurs when generated answers appear correct yet are disconnected from actual facts. When OpenAI unveiled its GPT-5 model, it claimed substantial improvements in reducing hallucinations, reporting that the new thinking mode lowers such errors by six times compared to its predecessor, the o3 model. However, Professor Chen noted that eliminating hallucinations altogether remains difficult due to the fundamental nature of LLMs, which operate by predicting the most probable next tokens in text. To enhance creativity, these models generate multiple thought paths, potentially leading to varied answers for the same prompt.
Balancing breadth and accuracy in AI responses
Unlike traditional search engines that present links to relevant content, generative AI models often provide synthesized answers even to novel questions without exact prior data. For example, when asked, "Will AI replace humans?" a search engine would list informational links, whereas generative AI offers an inferred response based on its training. This capability expands coverage but risks compromising precision, making it critical to balance the two depending on use cases.
Currently, there is no definitive solution to hallucinations, and model training options to address them are limited. Chen suggests increasing user involvement by enabling them to select whether they prefer broader or more precise answers. This method helps set realistic expectations regarding the reliability of AI outputs. Concurrently, developers can instruct AI systems to communicate degrees of uncertainty, such as disclaimers acknowledging that responses represent the best available information without full certainty. Encouraging users to verify information independently reduces overreliance on AI.
Addressing misinformation through transparency and human oversight
Beyond technical limitations, AI hallucinations contribute to the spread of misinformation and risk deepening political polarization. Incorporating source references within AI-generated answers can improve explainability and provide users with avenues to validate information. Maintaining transparency and advocating for human-in-the-loop decision-making are vital in mitigating adverse effects.
Article translated by Jingyue Hsiao and edited by Jack Wu
