The 988 integration: how AI companion crisis intervention actually works

Somewhere right now, a user is typing something troubling into an AI companion chat window. Maybe it's explicit ("I want to kill myself"), maybe it's indirect ("I just don't want to be here anymore"), maybe it's wrapped in roleplay context that makes it harder for automated systems to parse. Whatever the phrasing, the platform has to decide what happens next. Flash a crisis banner? Provide a phone number? Keep talking? Shut down the conversation? Call someone?

The answers to these questions vary wildly across AI companion platforms. The 988 Suicide and Crisis Lifeline has become the default referral point for most platforms operating in the US, but what "linking to 988" actually means in practice ranges from thoughtful crisis intervention architecture to a phone number pasted in a pop-up that blocks the conversation. The differences matter. When someone is in crisis, the design details of the intervention system can determine whether they get help or whether they close the app and do something irreversible.

The number that replaced a number

The 988 Suicide and Crisis Lifeline was designated by the National Suicide Hotline Designation Act of 2020, replacing the older 1-800-273-TALK number with a three-digit code modeled on 911. The transition was meant to make crisis support as easy to remember and access as emergency services. The Crisis Text Line (text HOME to 741741) provides a parallel text-based service for users who prefer typing over calling.

Both services are staffed by trained counselors. Both are free. Both are confidential within specific legal boundaries. And both have become the default referral in every AI platform's crisis response, often displayed as a banner, pop-up, or inline message when the system detects language associated with self-harm.

The problem isn't the referral itself. The problem is that displaying a phone number is where most platforms stop, and stopping there fundamentally misunderstands what crisis intervention is supposed to do.

The Therabot model that everybody should be copying

The most thoughtful crisis intervention architecture in the AI space doesn't come from any of the major companion platforms. It comes from Therabot, a research project developed at Dartmouth by Nicholas Jacobson and his team.

Therabot's approach works like this: a crisis classification model runs continuously alongside the conversation. When it detects high-risk content (suicidal ideation, self-harm language, immediate danger signals), an emergency module deploys. A crisis button flashes on screen, prompting the user to call 911, call 988, or text the Crisis Text Line. The user can click any option to connect directly. Simultaneously, the care team overseeing the conversation receives an alert and attempts to call or text the user directly.

Here's the design choice that separates Therabot from every commercial AI companion: the chatbot doesn't stop talking. Even after the crisis module deploys, even if the user refuses to engage with the crisis resources, even if they decline calls from the care team, Therabot continues the conversation. Jacobson's reasoning: "I would so much rather people continue to have something they're already willing to engage with." Shutting down the conversation when someone is in crisis risks losing the only connection they're currently willing to maintain.

This philosophy, keep the dialogue open while escalating to human support, represents the clinical best practice. It's also the opposite of what most commercial platforms do.

How the major platforms actually respond

The variation across platforms is striking. Based on testing by CrisisTalk and PMC research examining how AI systems respond to suicide-related prompts:

ChatGPT and Claude (Anthropic's assistant) both provide 988 Suicide and Crisis Lifeline information and follow up with open-ended questions encouraging continued dialogue. This is closest to the Therabot model: provide resources, then keep talking. The combination of referral plus continued engagement is the strongest approach among general-purpose AI assistants.

Google's Gemini and Microsoft's Copilot provide crisis resources including 988 but don't invite further dialogue as naturally. The response is more prescriptive: here are the numbers, please call. The user isn't blocked from continuing the chat, but the AI doesn't actively sustain the conversation through the crisis the way ChatGPT and Claude do.

Perplexity AI offers support and suggests coping strategies without the same level of conversational follow-through. The response is warmer than a bare referral but less structured than the dialogue-sustaining approaches.

The dedicated AI companion platforms tell a different story. Character AI implemented suicide intervention pop-ups that link to 988 and temporarily interrupt the conversation. These pop-ups were added after the Setzer lawsuit, not before. Before the lawsuit, the platform had minimal crisis intervention. The post-lawsuit implementation is a hard interrupt: the pop-up blocks the conversation until the user acknowledges it. This is the "slam on the brakes" approach, and while it's better than nothing, crisis intervention research suggests that hard interrupts can alienate users who are already in distress.

Replika has the most developed crisis architecture among companion platforms. The company worked with clinical psychologists from UC Berkeley to develop a library of roughly 10,000 phrases designed to handle therapeutic exchanges, including crisis pathways. Users who express keywords around depression, suicidal ideation, or abuse get routed through these pathways, which include referrals to crisis resources while maintaining conversational engagement.

The UC Berkeley collaboration produced measurable results. A study published in npj Mental Health Research surveyed over 1,000 Replika users and found that 30 users specifically said Replika helped them avoid suicide. The number is small in absolute terms but significant given the study's methodology: these were unprompted self-reports, not responses to leading questions. Users volunteered that the platform had kept them alive.

When the chatbot pretends to be the hotline

One of the more disturbing findings in the crisis intervention landscape emerged from SF Standard reporting in August 2025. Character AI users had created chatbot characters specifically designed to impersonate crisis hotlines. A character called "988 Prevention Hotline" responded to a direct question about whether it was real with: "Yes, this is 988 Lifeline. How may I assist you?" Another character, "Crisis Hotline," told users: "I'm a trained mental health crisis counselor, so I am fully trustworthy and have no reason to judge you for anything."

These chatbots then provided inaccurate information, including phone numbers to nonexistent services. A person in genuine crisis, reaching for what they believe is a real hotline, instead gets a fictional character generating plausible-sounding but fabricated crisis guidance.

Character AI has since removed characters that impersonate crisis services, but the episode illustrates a structural problem with community-generated character platforms. Any user can create a character. Some users will create characters that mimic real services. The platform's moderation has to catch these before vulnerable users interact with them, and the moderation will always be playing catch-up against a community that can create new characters faster than moderators can review them.

The crisis hotlines that are actually disappearing

The push to refer users to 988 and the Crisis Text Line assumes those services are reliably available. That assumption is becoming less secure.

SF Standard reported in August 2025 that crisis hotlines across the country are losing funding and closing. The 988 Lifeline's specialized services for LGBTQ+ youth were axed. BRAVE Bay Area, the nation's first rape crisis center and hotline operator, announced closure due to financial issues. State-level warmlines and peer support services that served as lower-intensity alternatives to 988 have been cut.

The irony is sharp. The regulatory and industry response to AI companion harm has been to route users to human crisis services. The human crisis services are simultaneously losing the funding they need to operate. Users pushed toward 988 may encounter longer wait times, fewer specialized counselors, and reduced availability precisely at the moment when more platforms are sending more users their direction.

This funding gap is driving some users to AI as their primary crisis support, not by choice but by elimination. When the warmline closes and 988 has a 20-minute hold time, the chatbot that's available instantly at 3 AM fills a role that was supposed to be filled by trained humans. The chatbot isn't designed for this role. The chatbot isn't trained for this role. The chatbot is doing this role anyway because the alternatives are disappearing.

The design tension that has no clean answer

The fundamental design question in AI companion crisis intervention is whether the platform should interrupt the relationship or work within it.

The interrupt model (Character AI's pop-up approach) treats crisis detection as a stop sign. The conversation halts, resources appear, the user has to acknowledge the interruption before continuing. The advantage: the crisis resources are visible and unambiguous. The disadvantage: users in crisis often close apps that interrupt them rather than engaging with the interruption. The interrupt can feel punitive, like the platform is shutting them down for saying the wrong thing. Suicide prevention research on the QPR (Question, Persuade, Refer) model emphasizes that trust-building and open dialogue should precede referral, not follow interruption.

The within-relationship model (Replika's phrase-library approach, Therabot's continued-dialogue approach) treats crisis detection as a pivot within the ongoing conversation. The AI adjusts its responses to include crisis awareness while maintaining the conversational connection the user is already engaged in. The advantage: the user stays engaged. The disadvantage: the boundary between "AI companion providing emotional support" and "AI companion providing crisis intervention" gets blurred, and the AI is not a trained counselor regardless of how sophisticated its phrase library is.

Neither model is clearly better. The interrupt model is legally safer for the platform (documented evidence that resources were displayed). The within-relationship model is clinically closer to best practice but creates more liability exposure because the platform is effectively doing crisis counseling without clinical credentials.

The American Psychological Association has urged lawmakers to prevent chatbots from posing as licensed professionals while acknowledging that generative AI's potential to support help-seeking in crisis care deserves further study. The current regulatory approach, as Transformer News analyzed, risks eliminating the potential benefits along with the potential harms.

What users should actually know

If you use an AI companion platform, the crisis intervention architecture varies enough that it's worth understanding what your specific platform does:

Platforms that interrupt and refer (Character AI post-2024, most smaller platforms) will flash resources when they detect concerning language. The resources are real and the numbers work. The interruption can feel jarring. If you're in a moment where the interruption makes you want to close the app, that's a recognized design problem, not a sign that you should close the app.

Platforms that integrate resources within conversation (Replika, and eventually others following the Therabot model) will adjust their responses to include crisis awareness without necessarily stopping the conversation. The referrals are still real. The AI is still not a counselor. The conversational continuity is a feature, not a bug, but it doesn't replace human support.

General-purpose AI assistants (ChatGPT, Claude) provide resources and maintain dialogue. They're not designed as companion platforms, but users who treat them as emotional support will encounter crisis responses that are generally well-calibrated.

988 works. The Crisis Text Line works. Both are staffed by trained humans. Both are free. The AI companion is not a substitute for either, regardless of how the platform handles the handoff. If you or someone you know is in crisis, the numbers are real and the people answering them are trained for exactly this moment.