Detecting risk isn’t enough for safer mental health AI

September 16, 2026

People in crisis are among the millions now turning to AI for mental health advice.

General-purpose AI tools are getting better at detecting risk and encouraging outreach to emergency services, which is progress in the right direction. At the same time, when someone is in danger, those steps are often not enough. The support available after risk is detected, and how well it matches the level of need, can be critical to whether people in crisis actually connect with care.

Risk can take many forms

Risk can show up in people’s lives in different ways, from thoughts of suicide or self-harm to threats of violence, psychosis symptoms, substance use concerns, or relationship violence. The signals aren't always obvious, either. People may describe what they’re experiencing in language that’s explicit, subtle, or ambiguous.

The next generation of Wellbeing AI needs risk assessment and management systems that can account for this complexity and connect people with qualified and specialized human support.

Recognizing risk doesn't resolve it

When general-purpose large language models identify certain signs of risk, they may provide a crisis hotline or encourage someone to seek emergency help. While those resources can be valuable, taking the next step to get the needed support and care is still largely up to the person in distress.

Someone who’s already overwhelmed and vulnerable can face a daunting process just to figure out where to turn, make the call, re-explain what they’re going through, navigate transfers or hold times, and choose the care they need. Even after reaching someone, they may face another hurdle: finding a qualified mental health provider who can see them quickly and provide the right level of support.

Mental health AI requires effective safety responses to risk

For risk management to be effective and have lasting impact we need to go further than detection. It should assess the nature, severity and urgency of the risk and connect the person with the human care they need.

A strong system for addressing risk in mental health AI includes:

Detecting a broad range of risk - AI recognizes potential risk and the many ways people may communicate risk across a wide spectrum of safety concerns, including suicide ideation, self-harm, violence, or homicidality concerns.

Responding quickly with human support - Risk detection triggers immediate outreach from a crisis support specialist who connects with the person, evaluates the situation, and determines the appropriate support.

Connecting people to the right level of care - Clinical support matching the level of need is available, with openings to see a provider within days, not weeks.

Continuing support after the first connection - Support is offered between the first outreach and the first care appointment, as well as ways to reconnect with the care team if concerns return.

Building risk management into wellbeing AI

Lyra combines AI risk detection across a wide range of safety concerns with prompt, clinician-led risk assessment and management. Clinicians provide additional care and interventions when wellbeing support is not sufficient.

When risk is detected, our Care Navigators reach out, often within minutes, to complete a risk assessment and determine the next step. They help members quickly connect with therapists and other resources when needed, rather than leaving them to find care on their own. Members can also get support from their providers and broader care team between sessions.

A higher bar for wellbeing AI safety

AI has tremendous potential to expand access to high-quality mental health support. Delivering on that potential responsibly requires managing risk at scale effectively.

For organizations evaluating wellbeing AI, it’s important to consider not only whether the tools can detect safety risks, but also if the system can respond effectively when someone is in danger. That deeper evaluation requires looking beyond the AI interactions to the care system around them: specialists who can step in, pathways to appropriate care, and support that follows.

To be there at the right time, in the right way, and with the right care is the reason Lyra was created. As we bring AI into mental health support, that commitment remains at the center of everything we build.