On November 18, 2025, the US House Energy and Commerce Subcommittee held a hearing on the risks and benefits of AI chatbots in mental health care. Experts including Dr. John Torous, director of digital psychiatry at Beth Israel Deaconess Medical School, shared insights on these rapidly emerging tools. In this interview, Dr. Torous reflects on the future of AI in mental health and the role of frameworks like MindBench.ai in shaping safe, patient-centered innovation.
AI at an Inflection Point
AI tools for mental health carry enormous promise—but, as Dr. Torous puts it, they are “feasible but fragile.” Their potential to transform care depends on robust regulation, rigorous research, and close collaboration between clinicians, patients, researchers, and technology developers. This is a pivotal moment: how stakeholders act now will shape whether AI fulfills its promise or falters.
Shifting Incentives for Safer AI
Dr. Torous stresses the importance of moving beyond metrics that reward engagement alone. Instead, AI tools should compete on safety, privacy, and clinical effectiveness. Lessons from the early mental health app market show that chasing clicks does not improve outcomes and can even erode trust. Establishing clear standards and incentives is essential to foster innovation that truly benefits patients.
Collaboration, Transparency, and Research
Developing effective AI tools requires a collaborative approach. Transparent, well-designed research helps establish patterns of use, potential benefits, and risks, while encouraging responsible data sharing. Collaboration across the ecosystem ensures that AI tools are both safe and evidence-based, and that improvements are shared as part of a collective effort rather than siloed in proprietary systems.
MindBench.ai: Patient-Centered Evaluation
At the heart of Dr. Torous’s vision is MindBench.ai, a framework currently in development that uses patient-informed benchmarks to evaluate AI mental health tools. MindBench is being co-designed with patient groups to ensure assessments reflect real-world experiences. It helps patients, clinicians, and developers make informed decisions, guiding responsible innovation and supporting the creation of tools that genuinely improve mental health outcomes.
Beyond Traditional Models
The next generation of AI tools has the potential to go far beyond replicating existing therapies. By integrating personal data on emotions, physiology, and environment, AI could inform entirely new approaches to care and even enable personalized treatments. Unlocking this potential, however, requires trust, thoughtful oversight, and careful handling of sensitive data.
Feasible but Fragile
AI in mental health is at a turning point. While the promise is immense, its trajectory remains fragile. With careful regulation, rigorous research, and patient-centered frameworks like MindBench, AI can evolve into safe, effective, and equitable support for mental health care worldwide.
Read the original article on the JMIR website here.
