Beyond Borders: Global Mental Health Conference

AI can play a supplementary role in mental health care rather than a replacement. Waitlists, last-minute cancellations, insurance barriers, and geographic shortages make conventional care structurally unavailable to much of the population it claims to serve. Yet current AIs hold assumptions about vulnerable users that actively misrepresent them.
This presentation identifies three distinct misreadings revealed in AI Chatbots, drawing on interdisciplinary evidence across human-computer interaction, clinical psychology, critical theory, and platform studies. First, AI outputs perform helpfulness and safety rather than faithfully tracking users' actual states, resulting in sycophancy, gaslighting, and over-sensitive crisis detection. AIs restructures the therapeutic space itself: users cannot know which disclosure will trigger a refusal or escalation, and so begin self-censoring by performing wellness rather than expressing distress. This dynamic is compounded by the growing solicitation of AI note-taking systems in clinical settings, where the presence of a monitoring apparatus changes the room before it acts. Second, AI tools are designed as products but experienced as relationships. Paywalls and abrupt model changes function not as consumer inconveniences but as structural instability and relational loss for the most dependent users. Memory failures further misrepresent users' narratives through premature theme-locking, stereotype overwriting, and detail fixation, treating people as types to be categorized, rather than complex entities evolving and unfolding over time. Third, the current therapeutic frameworks are predominantly built on a Western, individualist, and CBT-derived basis, which both AI and human therapists are prone to. Yet unlike human clinicians constrained by their training, AI holds genuine capacity for cross-cultural adaptability, which will support culture-specific care.
For AI and mental health wellbeing, it is not technical failures awaiting better engineering that matter most. It is the wrong assumptions about what vulnerable people need. Resilience cannot be built on a foundation that systematically misrepresents the person it claims to support.