Grow Therapy, Stanford Set AI Safety Standards for Mental Health

NEW YORK, June 23, 2026 — Grow Therapy (Grow) today announced a research partnership with Stanford University to develop rigorous, evidence-based standards for the safe use of artificial intelligence (AI) in mental health care. The collaboration is led by Principal Investigator Jonathan Chen, MD, Associate Professor of Medicine at Stanford, in coordination with Stanford Psychiatry.

The first study in this partnership will systematically measure how leading AI models respond to mental health crises and identify which design choices most effectively reduce harm while providing appropriate support. Researchers will analyze de-identified clinical cases that reflect real client safety concerns. These cases are drawn from conversations routed for human review by the screening systems of Grow’s AI coach, a conversational tool clients use between therapy sessions.

The study will combine automated evaluation with clinical review. A multidisciplinary research team, including a panel of psychiatrists and other mental health experts, will:

1. Measure how frequently prominent AI models produce responses that could be harmful or inappropriate in sensitive mental health situations;
2. Test whether adding a dedicated safety reviewer layer or similar checks meaningfully reduces harmful outputs; and
3. Assess additional safety mechanisms that, from mental health specialists’ perspectives, lead to more responsible and clinically appropriate responses.

By validating safety strategies against real-world scenarios rather than only synthetic examples, this independent effort aims to create a shared framework for evaluating large language models (LLMs) in the context of mental health conversations. The framework will help define what constitutes acceptable model behavior and make it easier for clinicians, researchers, and product teams to compare safety outcomes across different systems.

“People now regularly talk with LLMs, seeking therapy, advice, and counseling in both official and unofficial ways. To have an agreed upon definition of acceptable AI model behavior while mitigating predictable harm, we need studies like this to identify and measure failure modes. This will continue to push the field to hold high standards when patient care is involved.”

Jonathan Chen, MD, Associate Professor of Medicine at Stanford and Principal Investigator

Why Grow

Grow’s AI coach was developed by mental health professionals to support clients between therapy sessions, and the system was designed with multiple safety layers from the start. Each conversation is automatically scored for clinical quality, tone, and risk; interactions that trigger risk flags are routed to licensed clinicians for review. Grow’s approach is also informed by an external AI advisory panel composed of independent experts across clinical psychology, clinical AI research, ethics, and mental health epidemiology.

“It is essential that the mental health field develop evidence-based benchmarks for safe AI responses,” said Manoj Kanagaraj, MD, Chief Strategy Officer and Co-Founder of Grow Therapy. “We believe the strongest model keeps a licensed therapist in the loop, as ours does, but we recognize that some people may turn to general-purpose LLMs for mental health support. Partnering with Dr. Chen and his team at Stanford will establish clear standards, leveraging Grow’s real-world experience, to create a shared yardstick for AI safety in mental health. Our hope is that this work results in fewer harms reaching the people who turn to LLMs of any kind for mental health aid, especially general-purpose ones not designed for it.”

The collaboration emphasizes real clinical data over hypothetical examples. While many safety evaluations rely on synthetic testing or contrived prompts, this study uses scenarios that reflect the types of interactions clinicians actually encounter and worry about—particularly situations where an incorrect response could increase risk or cause harm.

“Much of the current testing of AI safety relies on hypothetical examples. This study gives us a chance to evaluate scenarios that are closer to what clinicians actually worry about in practice, especially in situations where the wrong response could cause harm.”

Ethan Goh, Executive Director of Stanford-Harvard ARISE

The research will produce measurable benchmarks and methodologies that other teams can adopt to evaluate LLM safety in mental health contexts. Expected benefits include clearer standards for acceptable model responses, evidence on which mitigation strategies are most effective, and guidance for product teams building conversational tools intended for sensitive use cases. Ultimately, the goal is to reduce the likelihood that people seeking support from AI-powered systems encounter unsafe or clinically inappropriate guidance.

Grow and Stanford will maintain patient privacy and confidentiality throughout the study by using de-identified material and following applicable ethical and regulatory standards for research involving clinical data. The study’s findings are intended to inform the broader mental health and AI communities and to encourage the development of robust safety practices wherever LLMs are used for support or advice.

Media Contact

Kristina McPherson [email protected]