Memory alone is not enough: Evaluating timing and omissions in long-term AI mental health support

Researchers, in collaboration with mental health professionals, developed the CARE (Context-Aware and Readiness-Evaluative) framework to evaluate the timing, quality, and omissions of advice in AI mental health support. Based on chatbot histories donated by 22 users, the study ran 14,256 longitudinal simulations to examine how an assistant equipped with memory responds across recovery trajectories and communication styles during suicidal ideation, self-harm, and anxiety crises. Results show that after a crisis, the match between advice and user readiness was lower than in matched control segments within the same conversation, mainly due to premature advice, which was most common when recovery stalled; revisiting a crisis reduced premature advice but did not reduce delayed advice; user communication style changed the frequency and timing of advice given. Clearly unhelpful advice was rare, but the assistant often omitted support that clinicians expected, such as follow-up and next-step planning. The authors note that these findings indicate limitations of safety evaluation at the single-response level, and that evaluating timing, omissions, and cross-user consistency is needed when holding AI systems accountable for long-term support.