When AI Corrects Users: How Communication Design Shapes Autonomy and Social Experience in Human–AI Interaction

Generative AI agents increasingly do more than provide information; they may also proactively correct users’ inaccurate statements. Such correction can improve informational accuracy but also create interpersonal sensitivity, making communication design and task context consequential for user experience. Drawing on Social Information Processing theory and politeness theory, this study examines linguistic politeness and correction style (supportive vs. defensive) across task-risk contexts. A mixed-design experiment (N = 380) showed that the high-politeness, supportive-correction configuration was associated with the most favorable perceived autonomy and social interactivity. Under higher-risk conditions, supportive correction attenuated the unfavorable association of low politeness with both outcomes, whereas defensive correction became less favorable across politeness levels. Bayesian multilevel mediation showed distinct pathways to service evaluation through autonomy and social interactivity, with credible indirect effects through social interactivity across all conditions. The findings distinguish politeness from correction style and provide a context-sensitive account of AI correction.