User vulnerability in artificial intelligence-based mental health interventions: Personality and developmental perspectives

Abstract

The rapid development of large language models is transforming psychiatric care, in part through the rise of chatbot-based digital mental health interventions. These artificial intelligence (AI)-driven interfaces function as a “digital first-aid kit” for individuals experiencing mental health challenges. Offering anonymity, immediate 24/7 access, personalized interaction, and low cost, they are often used before any contact with a human clinician. By lowering barriers to care, these tools have unprecedented potential to help close the global mental health gap. At the same time, they carry inherent risks. As Opel and Breakspear noted,¹ generative AI is susceptible to hallucinations and may align excessively with user intent, generating misguided or unsafe guidance that, in extreme cases, could contribute to suicide risk. While hallucinations and other model-level failures remain important concerns and may be partially mitigated through domain-specific fine-tuning and reinforcement learning,² they represent only one side of AI-related risk. The safety of AI-mediated mental health support may also be shaped by user vulnerability, including how individuals interpret, trust, and respond to AI-generated outputs. Drawing on the biopsychosocial model, we propose that personality characteristics, developmental processes, clinical vulnerabilities, and social contexts may shape these responses (Figure 1).

In general, generative AI produces responses that are linguistically plausible rather than clinically grounded. Individuals frequently engage with chatbots as trusted peers, relying on emotional subtext and culturally grounded shorthand that current AI interfaces cannot reliably interpret. Particularly in those seeking emotional reassurance, this interaction may foster a perception of empathy and understanding that exceeds the actual capabilities of AI systems. Such perceptions may increase the likelihood of anthropomorphizing chatbots and relying on their guidance in emotionally salient situations. It is possible for individuals with attachment insecurity, interpersonal sensitivity, loneliness, reassurance-seeking tendencies, or dependency-related personality traits to be more likely to perceive AI as trusted social partners. This tendency may be particularly relevant in individuals with mental disorders; conditions frequently associated with interpersonal sensitivity and personality-related vulnerabilities. Across these disorders, altered functional networks involved in self-referential processing, salience detection, and cognitive control may increase susceptibility to external guidance, although direct evidence linking these neural alterations to AI overreliance remains limited. Within this context, reliance on AI advice may be shaped less by model accuracy alone but by user characteristics that influence trust, interpretation, and dependence on AI-generated guidance. Individuals with high suggestibility or acute distress may be particularly vulnerable to developing an overreliance on AI generated guidance, potentially creating a “trust trap” in which ambiguous responses are interpreted as clinical truth. Such reliance may undermine adaptive coping strategies that are important for mental health resilience, including perspective shifting, mindfulness, interoceptive awareness, sleep regulation, and reflective practices such as journaling. Continuous access to immediate AI feedback may reduce reliance on self-regulatory mechanisms, fostering maladaptive coping centered on external rather than internal regulation. Chatbots may also lack the ability to guide users to reflect on past experiences, process trauma, or manage coping skills as therapists do. As a result, they may inadvertently displace therapeutic mechanisms that foster emotional integration and adaptive coping, making users feel interpersonal disappointment. One possible consequence is that repeated disappointing interactions with chatbots may weaken trust in subsequent therapeutic relationships with human clinicians and potentially influence future help-seeking behavior and treatment adherence. This highlights the importance of embedding AI tools within broader systems of care rather than positioning them as standalone substitutes for therapeutic relationships.

In addition, Opel and Breakspear¹ indicated that adolescents are a particularly vulnerable group in AI-mediated mental health support. However, emerging evidence suggests that socioemotional intelligence continues to mature extending into the mid-30s³ Neurodevelopment remains ongoing, particularly in prefrontal systems involved in executive function, emotional regulation, and critical evaluation. This may limit the capacity to assess the accuracy or ambiguity of AI-generated advice. Meanwhile, socioemotional intelligence and personality trait are not yet fully consolidated across this extended developmental window; core aspects of identity, values, and interpersonal attachment patterns remain in formation, increasing sensitivity to seemingly empathic or authoritative AI tools.⁴ From this perspective, concerns regarding overreliance on AI-delivered psychological support may extend beyond adolescence to the mid-30s. One theoretical possibility is that repeated internalization of AI-generated interpretations may contribute to changes in self-concepts, regulate emotions, and interpersonal expectations. Consequently, AI-mediated mental health interactions in youth and early adulthood through the mid-30s is not only clinical but developmental. This underscores the need for caution when deploying such technologies in populations undergoing developmental transitions, particularly those with psychiatric vulnerabilities.

These considerations are consistent with the framework of the biopsychosocial model, wherein developmental, psychological, clinical, and social factors jointly influence reliance on AI-mediated support.⁵ Rather than functioning as standalone therapeutic agents, AI tools may be most beneficial when integrated into clinician-supervised care systems, where they can augment therapy through structured supports such as coaching, case management, and peer support.

Yuan Cao: Conceptualization; methodology; investigation; supervision; formal analysis; visualization; writing—original draft; project administration; funding acquisition. Paulo Lizano: Conceptualization; methodology; investigation; funding acquisition; writing—review & editing; project administration; supervision; formal analysis.

The authors declare no conflicts of interest.

Ethics approval was not required for this study.

Data sharing is not applicable to this article as no new data were created or analyzed in this study.