Exploring Design Contributions to Mental Health AI: Structuring Clinical Judgment Through Real-Time Data Labeling Interfaces

This study explores how design can contribute to collecting training data for AI models in complex clinical settings, particularly in the context of mental health counseling. Conventional labeling methods only record outcomes, failing to capture moments of expert judgment. This limits AI’s ability to learn the nuances of clinical reasoning. To address this issue, the study draws on the Interactive Machine Teaching (IMT) framework. In collaboration with clinicians, researchers co-designed and iteratively improved a real-time labeling interface. The interface was then refined and used in actual mental health consultations. Using qualitative methods, we examined how the interface influenced judgment and the structural aspects of data creation. Clinicians began to see the interface not just as a documentation tool, but also as a Data Labeling Collaborator that helped them articulate and organize their thinking. While the interface enhanced the consistency and depth of clinical sessions, concerns were raised about whether real-time labeling might disrupt the natural flow of clinician-patient interactions. Additionally, we found that non-verbal cues play a significant role in clinical judgment, highlighting opportunities for future design integration. This study suggests that design can improve the quality of AI training data and facilitate the cognitive processes underlying expert decision-making. This approach offers new methodological pathways for human-centered AI development.