Turning to AI? Exploring Cognitive Mechanisms of Support-Seeking Decision-Making in Depression and Anxiety

Abstract

As conversational AI tools become increasingly available for emotional and social support, understanding who turns to them has become an important question for mental health research. This choice likely reflects a combination of AI features—how AI provides social reward consistently and at minimal effort cost—and the cognitive and affective traits and biases of the individual: how they process social feedback and weigh the effort costs of social interaction. Here, we examine whether two cognitive mechanisms theoretically associated with depression, social belief updating and social effort sensitivity, help explain individual differences in preference for AI versus human support, and how these mechanisms relate to depression and anxiety symptom severity. We predicted that individuals who process social feedback with a negative bias, those high in social effort sensitivity, and those with more severe symptoms of depression and anxiety, would show a stronger preference for AI over human support. Young adults (18 to 25 years old) completed two novel behavioral decision-making tasks, a novel AI preference survey, and measures of depression and anxiety symptom severity (N = 208 for the Social Beliefs Task, N = 202 for the Social Effort Task). In the Social Beliefs Task, participants formed and updated beliefs about how much simulated partners liked them across repeated trials, yielding parameters for prior expectations, learning rate, emotional reactivity, and uncertainty. In the Social Effort Task, participants chose whether to invest more or less effort in making an Instagram post in exchange for more or fewer likes, yielding reward and effort sensitivity parameters. Learning rate and emotional reactivity during social belief updating emerged as the most consistent positive correlates of AI preference: individuals whose beliefs and mood were more strongly impacted by social feedback reported a stronger preference for AI across measures, suggesting that AI may appeal to these individuals as a more cognitively and affectively stable alternative to human interaction. Prior uncertainty about social evaluation was negatively associated with AI preference, opposite to our expectations. The hypothesized negativity bias in learning rate was not associated with AI preference, though it was positively associated with depression and anxiety symptom severity, consistent with prior literature. Contrary to predictions, higher effort sensitivity was associated with less frequent AI use and a greater likelihood of choosing a human therapy platform over an AI one. Together, these findings suggest that responsiveness to social feedback, both cognitive and affective, rather than negatively biased social learning, underlies preference for AI over human support, and that the cognitive profiles associated with depression and AI preference are partially distinct.