Abstract
<title>Abstract</title> <p>Background Although AI (Artificial Intelligence) is increasingly used to provide workplace feedback, little is known about how AI feedback comes to be perceived as social support by recipients. Social support theory has been developed almost exclusively in the context of human-to-human interaction, leaving open the question of how AI's communication characteristics shape the type of support recipients perceive. This study examined whether AI feedback valence—positive (focusing on strengths and successes) versus negative (focusing on areas for improvement and weaknesses)—is perceived as different types of social support. Methods A randomized controlled trial was conducted with 109 Japanese gig workers who were randomly assigned to receive either positive or negative feedback from GPT-4o-based chatbots. Participants discussed their own gig work experiences with the chatbot, which then provided feedback based on their accounts. They interacted with their assigned chatbot over two sessions one week apart, with perceived social support (emotional, informational, and instrumental) assessed immediately after each session and at a one-week follow-up. Results Positive feedback was perceived more strongly as emotional support than negative feedback, while negative feedback was perceived more strongly as informational support than positive feedback. The timing at which these differences emerged varied by support type: informational support effects were most pronounced immediately after the first session, whereas emotional support effects were largest at the one-week follow-up. No significant condition differences were observed for instrumental support. Conclusions The type of social support perceived from AI feedback can be systematically shaped by the design of its communication. Positive AI feedback functions as emotional support, while negative AI feedback functions as informational support, providing a foundation for designing AI feedback systems that align with workers' diverse needs. Trial registration: ISRCTN10031278 (https://doi.org/10.1186/ISRCTN10031278). Registered 8 May 2026. Retrospectively registered.</p>