Are AI hallucinations the same as error information? A qualitative exploration of hallucination tolerance and resistance strategies

What the paper says

Purpose This study aims to examine the phenomenon of “AI hallucinations” – instances when AI-generated information, although produced confidently and coherently, deviates from the user's intent, contains factual inaccuracies or is contextually inappropriate. It investigates how both end users and AI algorithm engineers perceive these hallucinations and distinguishes them from conventional errors. Design/methodology/approach This study conducted in-depth interviews with large language model (LLM) engineers and users (N = 29). The interview transcripts were analyzed using thematic analysis to explore perceptions of AI hallucinations. Findings The study finds that AI algorithm engineers view hallucinations as a normal phenomenon in the early stage of technology development, while users show varying tolerance depending on the context. Based on the degree of hallucination tolerance, the analysis builds the “intention–criterion” framework to clearly delineate the boundary between hallucinations and errors. This framework also facilitates the categorization of user responses into three strategic types: total resistance, compromise and concession, and proactive adaptation. Originality/value This study introduces a novel “intention–criterion” framework that distinguishes AI hallucinations from traditional errors, offering a fresh perspective on user tolerance and information adaptive behaviors in human–machine interaction. The insights not only enhance understanding of user behavior in human–machine communication but also offer practical guidance for optimizing LLM performance and establishing robust industry standards. Peer review The peer review history for this article is available at: https://publons.com/publon/10.1108/OIR-04-2025-0233