The inherent hallucination of generative artificial intelligence (GenAI) poses social risks. In human–AI interaction contexts, verification behaviour is key to controlling information risks, preventing the spread of false information and reducing information pollution. The purpose of this paper is to model the explicit process of this behaviour, i.e. how it occurs and what is the strategy.
The critical incident technique was adopted to collect qualitative data from 22 interview participants. They contributed 44 true incidents of verification behaviour in human–AI interaction contexts. Thematic analysis was conducted on the qualitative data to identify the key phases of the verification process and behavioural elements.
Verification behaviour process is composed of three phases. Before verification, interaction scenario and interaction tool provide the context of verification and induce user’s different motivation. During verification, the strategy consists of four contents (including information content, content-source consistency, source link and thought chain) and seven methods (including continued interaction with the AI, source links tracing, online information search, cross GenAI interaction, interpersonal verification, practical verification, search in archive materials). After verification, there are outcomes with information verification results, updated system cognition, emotion appraisal, behaviour outcome and further give rise to two types of impacts: interaction feedback and interaction adaptation.
Our results help enhance understanding of verification behaviour in human–AI interaction contexts from the perspective of the user, adding to existing knowledge. These findings also inform initiatives aimed at educating the public about verification, providing strategy for them to applicate. Besides, this study engenders useful implications for designing verification experience; the macro-process depicts the naturally occurring mechanisms for the design to follow.
