Mengting Tu | Searcharxiv

arXiv2026

Seeing Is Not Knowing: The Epistemic Limits of LLM-Based Mental Health Inference from Youth Social Media Conversations

Adolescents increasingly use social media for private conversations, sharing experiences and seeking support. These interactions generate digital traces, prompting efforts to infer mental health and psychosocial states. As large language models become more capable of reasoning about conversations and interpersonal dynamics, a fundamental question remains nonetheless: what can models actually "know" about an individual's psychological experience from what they are able to "see"? We evaluated ten reasoning LLMs using donated Instagram Direct Messages from 80 adolescents with a history of depression or suicidal thoughts and behaviors, comparing predictions with self-reports and analyzing 204 reasoning traces. Models performed better when inferring more observable states than participants' subjective appraisals of social experience, such as perceived social support. Beyond prediction accuracy, our findings highlight the epistemic limits of digital traces for AI design and evaluation: what can be observed, what can reasonably be inferred, and what requires people's own accounts.

cs.HC↗

arXiv2026

Remembering Is Not Enough: Evaluating Timing and Omission in Longitudinal AI Mental Health Support

Evaluations of large language models (LLMs) for mental health support commonly focus on the clinical appropriateness of individual responses, potentially overlooking failures that emerge across extended interactions. As AI assistants accumulate memory of users' circumstances and prior disclosures, this raises an accountability question: Do AI chatbots provide appropriate support as users' needs evolve? We investigate by analyzing 14,256 longitudinal simulations grounded in donated chatbot histories from 22 users, examining how a memory-equipped assistant responds to suicidal ideation, self-harm, and anxiety crises across recovery trajectories and communication styles. We introduce the Context-Aware and Readiness-Evaluative (CARE) framework, developed with mental health professionals, to assess the timing and quality of advice, including omissions. Advice was less aligned with users' readiness after a crisis than in matched control stretches of the same conversation, largely because of premature recommendations, which were most common when recovery stalled. Revisiting a crisis reduced premature advice but not lagging advice, and users' communication style changed how often and when the assistant gave recommendations. Although explicitly unproductive advice was rare, the assistant frequently omitted clinician-expected support such as follow-up and planning for next steps. These findings show the limits of single response-level safety evaluation and the need to assess timing, omission, and consistency across users when holding AI systems accountable for longitudinal support.

cs.CY↗