By John Draper, Ph.D.
Part II of a three-part series on AI in crisis care. Read Part I, which made the case for keeping human connection at the center of crisis care. Part III will examine how AI can support the crisis workforce.
Nearly one in five Americans ages 12–21 reported using an AI chatbot for mental health advice in the past year, according to a 2026 JAMA Pediatrics national survey; most had not told anyone they were doing so. OpenAI has separately estimated that about 0.15% of its weekly active users have conversations containing explicit indicators of potential suicidal planning or intent. At this scale, even a small percentage represents many people.
For some, an AI conversation may be the first place they disclose emotional pain. That creates an opportunity to help struggling people who might not otherwise contact a crisis line or seek mental health treatment.
In Part I, I made the case that authentic human connection must remain at the center of crisis care. But keeping human connection at the center of crisis care does not mean keeping AI at the margins. AI may help reach people who live where services are scarce, fear stigma or other consequences of seeking help, or are not ready to tell another person what they are experiencing.
The question is whether we design AI primarily to become the help—or to help people find their way toward it. I believe one of AI’s greatest opportunities lies in becoming the front door, coach and bridge to a much larger ecosystem of support.
Key Takeaways
- AI is becoming a first point of disclosure.
- In one randomized study, an enhanced digital crisis response increased self-reported use of offered resources from 39% to 78%; this did not verify connection to care.
- The goal is a menu of on-ramps, not one off-ramp.
Turn the Off-Ramp into an On-Ramp
Former NIMH Director Thomas Insel has described a problem with many digital approaches to suicide prevention. Someone may spend considerable time disclosing deeply personal distress to a chatbot, only to encounter what amounts to a “hard redirect” when suicide enters the conversation: Here is a crisis number.
From the standpoint of a safety protocol, that may make sense. From the standpoint of the person experiencing the crisis, it may feel like hitting a tripwire. Instead of ending the conversation with a referral, Insel argues for building a bridge of nudges and supports that helps move someone from the world of algorithms toward actual care.
I think that’s exactly right. Sometimes 988 is the right next step. But someone experiencing distress might instead need a domestic-violence service, peer support, psychoeducation about panic attacks, a brief evidence-based intervention, a story of recovery, or help finding ongoing mental health care. And they may have no idea which one they need.
The intelligence we should be building isn’t simply the ability to detect suicide risk. It is the ability to help determine what this person needs next—and help them get there.
Finding People We Currently Miss
People don’t necessarily begin a crisis by dialing 988. They Google symptoms. They post something concerning on social media. They search for information about self-harm. They tell a chatbot they haven’t gotten out of bed for three days. They ask whether what they’re experiencing is normal. If that’s where people are going, let’s meet them there and build systems that can accommodate their journey from wherever their preferred starting point happens to be.
Koko and ThroughLine illustrate different ways of creating this kind of digital front door. Koko has used machine learning in participating online environments to detect more than 1,300 crisis-related terms—including slang and deliberately disguised spellings—and offer users self-help, peer support and crisis resources. Its work is especially interesting because it has also focused on the barriers that prevent people from acting on a referral.
ThroughLine provides infrastructure connecting digital platforms with verified crisis and support resources around the world. Its newer FirstStep AI conversational guide is designed to understand a person’s needs through conversation and direct them toward an appropriate next step, including self-help resources or human support.
The front door doesn’t have to provide everything. But it does need to help people find the right pathway for them, beyond the door.
Give People Something Useful Right Now—and Build Agency
There is an enormous middle ground between “keep talking to the chatbot” and “call a crisis line.” There are brief interventions, educational resources and coping tools that can be useful before, during and after someone connects with care. And whether a person with suicidal thoughts is already receiving care or not, strengthening their sense of agency—their belief that they can take actions that make a difference—is critical.
Research with people who have experienced suicidal crises repeatedly points toward active coping: connecting with other people, positive self-talk and self-compassion, spirituality, peer support, grounding, keeping busy, restricting access to lethal means, and doing something purposeful. Research with suicidal Lifeline callers has similarly emphasized the importance of respecting people’s desire for self-efficacy and strengthening their capacity for self-care and informal support rather than assuming that professional treatment is always the only useful next step.
More recently, Thomas Joiner and colleagues have proposed an even more direct connection between suicidality and the loss of agency. Their 2025 Psychological Review paper, “The Death of Agency as the Spur of Suicide”, proposes that a profound erosion of one’s felt ability to affect what happens may contribute to the suicidal state. Importantly, restoring agency need not begin with a life-changing decision. It can begin with small purposeful actions that gradually restore a sense that what I do matters.
That has important implications for AI. Rather than creating systems that encourage people in crisis to become increasingly dependent on a chatbot, we should design AI to help people discover what they can do next: identify what helped them survive previous crises, practice grounding or cognitive reappraisal, develop a safety plan, choose one achievable action, or rehearse what they want to say to a family member, peer or counselor.
In other words, AI should not only provide answers. It can help people recover the capacity to act.
Single-Session Interventions: Help When Someone Is Ready for Help
One promising example comes from Jessica Schleider and colleagues’ work on Single-Session Interventions (SSIs)—brief, structured interventions designed to provide something useful in a single encounter rather than assuming that someone will begin and remain in traditional treatment. Some of this work has already been embedded in places where young people naturally seek help. In collaboration with Koko, Schleider and colleagues evaluated three 5–8 minute SSIs offered as Koko “minicourses” on Tumblr. Over a year, the interventions were viewed more than 18,800 times and completed more than 6,100 times, with users rating all three above 4 out of 5 stars. Most importantly, among those who completed the interventions, two were associated with significant reductions in hopelessness, while another produced significant reduction in self-hatred and increased participants’ desire to stop self-harming.
These aren’t intended to replace therapy or crisis lines. They demonstrate something important for the role I envision for AI: we don’t necessarily have to choose between leaving someone alone with their distress and immediately referring them somewhere else. Someone struggling at 2:00 a.m. may not be willing or able to enter psychotherapy or call a hotline at 2:01. But a brief intervention may still help reduce hopelessness, strengthen motivation to stop self-harm, teach a coping skill, or help the person identify one achievable next step. AI could potentially help make these interventions available at precisely those moments—and eventually help determine which kind of intervention is most likely to be useful to which person.
Newer research is also beginning to tell us what within these interventions may matter. A 2026 study by Ghosh, Kaveladze, Schleider and colleagues examined individual components of a behavioral-activation SSI. The study found concrete action planning had the clearest longer-term effect on depression, while psychoeducation reduced hopelessness immediately and testimonials increased feelings of relatedness. The individual effects were small, and the findings don’t yet tell us precisely which intervention should be given to which person. But the authors point toward an intriguing future: adaptive or “precision” digital interventions that deliver different components based on what someone needs.
That suggests a very different role for AI than generic “AI therapy.” Someone confused or frightened by what they’re experiencing might benefit from psychoeducation. Someone immobilized may need help constructing one achievable action. Someone feeling profoundly alone might benefit from peer support—or hearing how another person survived a similar crisis.
There is research behind that last possibility as well. Studies of the Papageno effect suggest that exposing persons in suicidal distress to carefully constructed stories showing how people coped with suicidal crises can reduce their suicidal ideation. Importantly, the evidence is about the effects of recovery-oriented narratives, not a license to show any suicide-related story to anyone. Randomized research with psychiatric patients, for example, found benefits among patients with affective disorders but not across every diagnostic group. AI might eventually help select appropriate stories of actual human lived experience rather than pretending to possess lived experience itself.
The intelligence we should be building isn’t simply the ability to detect suicide risk. It is the ability to help determine what this person needs next.
Bridging Referral to Connection
Perhaps the most compelling evidence for this “on-ramp” idea comes from a randomized study conducted by Koko and Schleider and colleagues. Social-media users identified as potentially being in crisis were directed to Koko and randomized to receive either standard crisis-line information or an approximately one-minute enhanced crisis-response SSI. The enhanced intervention included validation and psychoeducation, testimonials, a safety-planning option, brief evidence-based interventions, and crisis resources. When users said they were unlikely to use the resources, Koko provided an interactive intervention addressing barriers to doing so. The results were striking: 78% of participants receiving the enhanced intervention reported using a resource offered to them, compared with 39% receiving the standard crisis referral. They also reported significantly greater reductions in hopelessness ten minutes later.
That distinction matters. Koko didn’t make the crisis services themselves better. It made it more likely that people would take a step toward help. And the study did not verify that 78% connected with a human provider or establish longer-term suicide-prevention outcomes; resource use was self-reported shortly after the intervention. But as a demonstration of how a very brief digital interaction can potentially move someone from referral toward action, it is important.
This is where AI could become enormously useful. Instead of repeatedly telling someone to seek help, it can potentially help us understand why they aren’t seeking it, address that barrier, preserve their choices and agency, and make the next step easier.
A Menu of On-Ramps
Insel’s Benchmark Health offers another piece of this model, this time with humans explicitly in the loop. Benchmark assigns a human Advocate who helps a person understand options, find appropriate care, navigate scheduling and follow up on whether that care actually fits and is helping. Benchmark describes its approach as combining direct support, specialty-care navigation, care coordination, and longitudinal measurement. It is not itself an emergency crisis service, but the model illustrates something crisis systems badly need: not simply giving people referrals but helping them successfully navigate what comes next.
Taken together, these developments suggest that we shouldn’t build one universal AI off-ramp called “988.” We should build a menu of on-ramps: perhaps a brief SSI when someone needs hope and an achievable next step; psychoeducation when understanding can reduce fear; a carefully selected story of recovery; a coping or safety-planning tool; peer support when lived experience matters; a crisis counselor when immediate human support is needed; or human care navigation when someone needs help finding ongoing treatment.
What AI Suicide Prevention Means for Crisis Systems
Most of the innovation I’ve described is happening outside traditional crisis systems, on platforms where people already are. That creates both a gap and an opportunity for those of us who build and run crisis care.
- Treat digital front doors as potential partners. If people disclose distress to a chatbot or search engine first, those platforms can become important entry points into crisis care. Clear pathways between them and local services will matter more each year.
- Measure connection, not just referrals. The Koko study’s lesson is that how a referral is offered changes whether people act on it. Systems that track whether people actually reached help will learn far more than those that count referrals made.
- Make the next step easy to take. An on-ramp only works if something is waiting at the other end. Follow-up, care navigation and warm handoffs are what turn a moment of disclosure into ongoing support.
- Build AI into the system deliberately and safely. AI-enabled pathways need clinical governance, privacy safeguards, evaluated interventions, and reliable escalation to qualified human support when urgent risk is identified. As states plan their next generation of 988 and crisis infrastructure, the question isn’t only what AI should avoid, but how it can bring people toward effective human help.
The Front Door, Not the Destination
The intelligence isn’t merely recognizing distress. It is figuring out what might help this person, right now, offering meaningful choices, and making the next step easier to take.
That is the promise I see in AI as a front door, coach and bridge. The chatbot doesn’t have to become the destination. Its greatest value may be helping each person find the right on-ramp—and helping them take the first step onto it.
How is your crisis system thinking about digital front doors? Dr. Draper welcomes the conversation. Connect on LinkedIn or email jdraper@ihrcorp.com.
Additional Reading
For readers interested in going deeper, I recommend:
- Cohen, Dobias, Morris & Schleider’s randomized Koko crisis-resource study
- Dobias, Morris & Schleider’s study of SSIs embedded within Tumblr
- Ghosh et al.’s 2026 study of the active ingredients of digital SSIs
- Joiner et al.’s 2025 paper on suicide and the loss of agency
- JAMA Pediatrics national study of youth use of AI for mental-health advice
John Draper, Ph.D., is President of Research, Development & Government Solutions at Behavioral Health Link and founding Executive Director of the National Suicide Prevention Lifeline (now 988).