A New Role for AI in Mental Health: Support Before Crisis

There is a largely overlooked interval in mental health support: the period when resilience begins to weaken, loneliness deepens and emotional balance gradually shifts – but suffering is not yet visible or severe enough to prompt someone to seek help.

Imagine Maya, a college student sitting alone in her dorm room after midnight. Nothing dramatic has happened. She still attends classes, completes her assignments and answers, “I’m fine,” when someone asks how she is doing. Yet more messages remain unanswered. She declines invitations more often. At night, her worries become louder as the distance between her and other people quietly grows. Maya does not think she needs help. She is not in crisis – but she is no longer quite herself.

The future of AI in mental health may not begin with replacing the therapist. It may begin much earlier – helping people build resilience, connection and emotional awareness as difficulties first begin to emerge.

A six-week randomized trial published in NEJM AI tested whether an AI-powered well-being app could play this earlier role. Its findings do not prove that AI can prevent mental illness. But they suggest that mental health can be strengthened before professional treatment becomes necessary.

The gap before care begins

Mental health support frequently begins after symptoms become recognizable – when anxiety interferes with daily life, depression can no longer be hidden or distress reaches a crisis. Proactive support starts earlier. It does not claim to predict or prevent mental illness; it offers accessible help before someone must identify as ill or decide that their suffering is serious enough to justify care.

The need is substantial. In a national survey of more than 80,000 college students, 76% said they needed help for emotional or mental health problems, yet only 38% had received professional support during the previous year.

The NEJM AI study entered this gap. Researchers randomly assigned 486 students from three U.S. institutions either to receive access to Flourish or to continue with the support ordinarily available to them. Students were asked to use the app at least twice a week for six weeks, generally for only 5 to 20 minutes per week.

Compared with the control group, students given Flourish reported greater resilience and belonging, less loneliness and more positive emotions. They were also better protected against declines in mindfulness and overall flourishing. The study therefore asks whether carefully designed AI can help strengthen mental health before professional treatment becomes necessary.

AI as a bridge, not a destination

Flourish is an AI-powered well-being app built around Sunnie, a generative AI conversational coach. An interaction might begin with a brief emotional check-in. The AI then responds to what the student shares and recommends a small, evidence-based activity suited to the moment.

That activity may involve journaling, reframing an unhelpful thought, practicing mindfulness or identifying three good things. Other suggestions deliberately move beyond the screen: taking a walk, expressing gratitude, starting a meaningful conversation or seeking professional support when needed. Weekly summaries help users recognize patterns in their emotions and behavior.

This is what “carefully designed” means in practice. The conversations and activities were developed with psychologists and informed by established approaches from positive psychology, cognitive behavioral therapy and other behavioral sciences. The app was also designed to recognize high-risk language, initiate a safety check and direct users toward crisis or professional resources when necessary.

Unlike some AI companions that seek to prolong engagement, Flourish offers a healthier model: it uses the conversation to strengthen the person’s capacity to reengage with life, human relationships and professional support. The healthiest AI companion may not be the one that keeps us talking to the machine, but the one that helps us return to other people – and recognize when we need human help.

Well-being is not treatment

The trial did not show that AI can treat depression, anxiety or stress. None of those outcomes improved significantly. Nor did the study test whether the app could prevent self-harm, suicide or future mental illness.

The findings also came from a general population of college students over only six weeks. Outcomes were self-reported, the effects were modest and the control group did not receive an alternative app. We therefore cannot know how much of the benefit came specifically from generative AI rather than from structured reflection, positive activities or simply receiving additional support. Several authors also had financial ties to the company that developed Flourish.

The results are encouraging, but preliminary. AI should not replace therapists, crisis services or human relationships. Its more immediate promise may be to complement them – offering accessible support earlier, strengthening important capacities and helping people recognize when human care is needed.

Final thoughts

Resilience, connection and emotional balance can begin to weaken long before distress becomes a diagnosis or crisis. Support should not wait until then.

One six-week trial cannot tell us whether AI will prevent future illness. But it raises a more useful question: Can carefully designed AI help people remain connected to themselves and others before professional treatment becomes necessary?

The best mental-health AI may ultimately be measured not by how attached people become to it, but by how effectively it helps them return to life – and how reliably it recognizes when the next step must be human.