"This is My Fault", Really? Understanding Blind and Low-Vision People’s Perception of Hallucination in Large Vision Language Models

DOI

Visual question-answering (VQA) tools powered by large visual language models (LVLMs) are used to assist blind and low-vision (BLV) individuals in overcoming visual challenges, raising concerns about hallucinations and associated risks. Existing literature overlooks the variations of hallucinations across distinct usage scenarios and types in the context of VQA for BLV people, resulting in limited understanding of their perceptions and insufficient guidance for targeted mitigation strategies. By analyzing 3,467 real-world VQA cases from BLV users, we developed a manifestation-scenario-based dual-dimensional hallucination typology, uncovering eight scenarios and five types of hallucinations. Through interviews with 16 BLV users, we examined their awareness levels, detection strategies, mental models of hallucinations, and their tolerance of associated risks, identifying key gaps between their perceptions and real situations. By designing with 12 BLV users, we uncovered their expectations for hallucination-mitigating solutions, including enhanced information provision, transparency in processing, verification strategies, and feedback mechanisms.