RL mental health system reduces burnout, boosts journaling engagement

Researchers Tony Wang and Qian Yang from Cornell University explored whether reinforcement learning (RL) could bridge the gap between clinical mental healthcare and everyday wellness support. They developed a contextual bandit algorithm that dynamically selects journaling prompts from two repertoires—clinical (e.g., CBT exercises) and wellness (e.g., gratitude prompts)—to optimize for sustained journaling behavior. The system was deployed in a four-week exploratory study with 38 participants.

Key findings challenge conventional assumptions about intervention intensity. First, many benefits of RL-optimized sequences appeared only after the intervention period ended, suggesting AI systems should include “stepping-back” phases to consolidate gains. Second, participants who engaged most with RL-generated prompts deepened their engagement over time, while those receiving constant (non-adaptive) prompts showed higher burnout and dropout rates. This raises a critical design question: should clinical-wellness AI systems reduce intensity to prevent burnout, or maintain it to maximize treatment gains? The work points toward adaptive, person-centered digital mental health that respects natural waxing and waning of symptoms.