Impact of task type and explanation of failure reasons on users’ willingness to forgive in generative artificial intelligence: cognitive neural mechanisms

Abstract

Purpose By virtue of its own powerful generation ability, generative artificial intelligence (GenAI) has been widely used in many service scenarios. However, the situation of service failure still cannot be completely avoided. Currently, the exploration of remedy strategies that match the unique capabilities of GenAI is still in an insufficient state. Therefore, this study aims to explore the impact of different task types and failure attribution explanations on users’ forgiveness willingness in the context of GenAI. Design/Methodology/Approach This study integrates the associative-propositional evaluation (APE) model and expectancy disconfirmation theory (EDT), employing a multi-study approach with scenario-based experiments and event-related potential (ERP) technology. Study 1a (N = 480) utilized a 2 (task type: innovative vs. mechanical) × 2 (explanation of failure reasons: internal vs. external) × 2 (severity of failure: mild vs. severe) between-subjects scenario experiment to measure users’ deliberate forgiveness attitudes. In the general daily consumption scenario, Study 1b carried out relevant explorations (N = 287), using a between-subjects design of 2 (task types) × 2 (explanation of failure reasons) to test the stability of the core mechanism. Study 2 is an ERP experiment of 2 (task types) × 2 (explanation of failure reasons), capturing the key neural components N2 and P3, so as to track the automatic and implicit neural responses of users when facing AI explanations in real time. Findings Study 1 shows an extremely obvious three-way interaction among task type, failure severity and failure reason explanation, and this interaction is tightly dependent on specific situations. ERP data show that there is a two-stage neural processing flow: first, the cognitive conflict caused by “task expectation–reality” is quickly detected by the interaction of N2 amplitude; then the P3 component reflects that the brain will allocate more cognitive resources to deal with the higher uncertainty brought by external attribution. Practical implications This exploration has directly given experiential guidance to the design and interaction strategies of artificial intelligence products. AI service recovery systems must abandon the “one-size-fits-all” apology template and instead develop adaptive strategies that can dynamically adjust based on the nature of the task, the consequences of failure and the user’s history. Originality/Value For the first time, the APE model and EDT were integrated and applied to the failure scenarios of GenAI services, and the applicable boundaries of the traditional attribution theory in the field of human–computer interaction were corrected. In terms of methodology, it innovatively combines behavioral experiments with ERP neural experiments, providing cross-level and more powerful evidence for the “dual-path” psychological mechanism of user forgiveness than a single method.