Large language models identify but struggle to resolve therapy ruptures
Can LLMs identify and repair ruptures? Comparison between clinician practices and LLM behaviors
Human-Computer Interaction
Summary
Sometimes, conversations between a therapist and a patient encounter problems called ruptures, where trust or understanding breaks down. This paper looks at how well large language models (LLMs), a type of AI, notice these problems and try to fix them. The authors found that LLMs are good at spotting obvious problems from clear language clues but not as good as experts at understanding the deeper context. When trying to fix ruptures, LLMs tend to give scripted advice, while experts use more open and thoughtful responses. The study suggests improvements for AI systems used in mental health to be more sensitive and timely.
What this means in practice
- •For mental health technology developers: Improve conversational AI by focusing on relational context and timing to better handle trust breaks in therapy sessions.$Commercial implications: Enables more effective and engaging mental health chatbots that can better support users by recognizing and repairing conversational problems.
- •For customer service platform teams: Adapt AI tools to identify explicit conflict signals in conversations to flag potential breakdowns for human follow-up.
Authors
Jeongah Lee, Joy Qiuyue Zhong, Drishti Goel, Violeta J. Rodriguez, Dong Whi Yoo, Koustuv Saha, Ravi Karkar
Abstract
Ruptures represent common albeit critical moments in interaction where relational alignment breaks down, making them essential for evaluating AI where trust and engagement matter most. In a scenario-driven empirical study, we examined the performance of three LLMs at identifying and resolving ruptures across 21 mental health conversations and 22 experts' evaluation of the strategies. For identification, LLMs relied on explicit linguistic cues within single turns whereas experts integrated implicit, relational, and contextual information across the conversation. For resolution, LLMs tended to produce more directive and scripted responses whereas experts adopted process-oriented strategies such as validation, open-ended exploration, and psychoeducation. Overall, LLMs showed higher agreement with predefined labels in identification, but not in resolution where experts rated their responses only moderately effective, with consistent limitations in timing, depth, and contextual sensitivity. We discuss implications for the design of mental health conversational agents emphasizing relational awareness, pacing, and human-in-the-loop support.