【摘要】 针对目前“AI幻觉”研究多未回应AI技术已实现多模态内容生成的现实情况，研究梳理了“AI幻觉”的背景和内涵，提出了“AI幻觉”的新定义，并将其划分为形式幻觉、事实幻觉和混合幻觉三种主要类型。采用案例分析法，对“AI幻觉”在文字、图片、视频与音频等不同模态生成式AI平台中的典型表现进行了测试和说明。进一步，以各类生成式AI的Transformer架构为依据，总结出“AI幻觉”在数据集、数据处理方式和技术运行逻辑等方面的成因，并介绍了相应的应对措施。更多还原

【Abstract】 The study begins by highlighting that most of the current research on “AI hallucination” overlooks the reality that AI technologies have already enabled multi-modal content generation. It then offers a revised definition of the term. Building on this foundation, the study categorizes “AI hallucination” into three main types: modality hallucination, factuality hallucination, and hybrid hallucination. It employs case analyses across various modalities—text, image, video, and audio—to examine how hallucinations manifest in different forms. Furthermore, drawing on the Transformer architecture underlying various types of generative AI, the study explores the causes of AI hallucination from the perspectives of datasets, data processing methods, and technical operational logic, and proposes corresponding countermeasures.更多还原