AI-Enabled Deception Detection for Mental Health

An Exploratory Analysis of Automated Deception Detection for Mental Health Applications

Research overview

This dissertation combined qualitative research and multimodal machine-learning analysis to examine whether AI-enabled behavioral analysis could be developed and used responsibly in mental-health settings.

In interviews with mental-health professionals and six virtual focus groups with 18 participants who were both clinicians and current therapy clients, participants described concerns about false positives, model robustness, cultural variation in behavioral cues, confidentiality, client autonomy, and potential disruption to therapeutic relationships. Their perspectives informed design and policy recommendations centered on clinical judgment, client agency, and responsible data governance.

The computational work developed the Cross-Domain Deception Dataset (CD3), with truthful and deceptive video responses to biographical, academic, and well-being questions. Together with a publicly available crime dataset, CD3 enabled comparisons across four contexts with different stakes. The analyses found that behavioral features did not transfer uniformly: gaze and gesture could be more predictive than facial features, and the usefulness of cues varied with context.

This work contributes an empirical basis for asking when behavioral signals may transfer across settings, when context limits their interpretation, and when a model’s output should not guide a decision. It motivates my broader research program on context-sensitive behavioral modeling and human-centered AI systems that preserve human judgment and oversight.