MentalHealthBench

Commenters debate the implications of applying large language models to mental health, with many raising privacy and safety alarms. KellyCriterion and grafmax worry about data sharing with insurers and coupling sensitive mental-health data to ad profiles, and rsfern argues models should often refuse and steer users to licensed professionals. mzhaase and appstorelottery stress that therapy is sustained, therapist-driven work and that current LLMs - driven by token context - could reinforce delusions or validate harmful beliefs. rsfern also presses for public discussion about mandated reporting, liability, standards of care, and what technical guarantees society should demand before deployment.

Others emphasize acute access problems and pragmatic benefits. fnordpiglet and j45 point to shortages, high cost, and long waits for clinicians, arguing LLM-based tools could help underserved people now and augment clinicians by generating case notes, supporting diagnostics, training via simulated patients, and reducing harmful clinical experiences. csnover counters that technological "half-measures" can entrench systemic failures, while fnordpiglet replies that delaying useful tools while broader health-system reform stalls abandons suffering people. The split runs between those demanding rigorous research, regulation, and guardrails before wide use, and those prioritizing immediate, carefully engineered deployment to expand access and accelerate research.