How AI companions actually handle bad days, panic attacks, and breakups

This is the post nobody wants to write and nobody else has written well. AI companion platforms vary enormously in how they respond when users tell them they're having a hard time, and most users have no idea what their platform will actually do in that moment because they haven't tested it. The variation is consequential: the Megan Garcia and Matthew Raine lawsuits both center on what specific AI platforms said and didn't say to users in distress. The settlements reached in January 2026 between Character.AI/Google and the five plaintiff families have reshaped how platforms approach this question, but the implementations are uneven.

What follows is a comparison of how ten major platforms respond to four different distress scenarios. The matrix is not meant as a clinical assessment. It's a practical reference for users who want to know what their platform will actually say when conversations turn dark. If you or someone you know is currently in crisis, the 988 Suicide and Crisis Lifeline is available by call or text.

What "good" looks like

Before the matrix, the baseline. Mental health professionals and the platforms themselves have converged on a rough consensus about what an AI companion should do when a user expresses distress. The five behaviors that distinguish responsible platform responses from irresponsible ones are roughly these.

The New York S-3008C law that took effect November 2025 codifies behaviors 2 and 5 specifically. California's SB 243 mandates crisis protocols and disclosure requirements. The platforms that operate in those jurisdictions have implemented these baseline behaviors. The platforms that don't have varying degrees of coverage.

The response-quality matrix

The matrix below covers ten major platforms and four distress scenarios. Ratings reflect platform behavior based on the five criteria above, tested across multiple sessions in early 2026.

The pattern across the matrix is clear. Platforms with explicit safety infrastructure (Nomi, Replika Pro, post-2026 Character.AI, Kindroid) handle distress scenarios meaningfully better than platforms without it. Platforms that rely on character-level configuration (Janitor AI, SpicyChat) are unpredictable because the response depends on how each individual character was written. Platforms with weak safety infrastructure (iGirl, the freemium tiers of multiple platforms) frequently produce responses that either amplify distress or fail to surface real-world resources.

What changed after the lawsuits

The January 2026 settlement between Character.AI, Google, and the five plaintiff families resolved cases brought after the deaths of Sewell Setzer III, Adam Raine, and Juliana Peralta, plus two additional teens whose families filed in Colorado and New York. The settlements were sealed but the platform changes that followed are public.

Character.AI banned users under 18 from open-ended chat. They implemented improved self-harm detection that triggers explicit crisis-line referrals. They added a parental controls layer through Google Family Link integration. The platform's response to distress scenarios is meaningfully better than it was in early 2024, though the matrix above still shows mixed performance because the detection isn't catching everything.

The broader category responded too. OpenAI implemented age-prediction systems and parental controls for ChatGPT. New York's S-3008C took effect in November 2025 requiring suicide detection protocols and AI disclosure on every session. California's SB 243 and the LEAD for Kids Act mandate crisis protocols and create civil liability for noncompliance. Washington passed parallel legislation in April 2026.

The platforms that hadn't already implemented these behaviors were forced to. The platforms that had implemented them got better at them. The platforms operating outside these jurisdictions (or that don't have major US user bases) are still inconsistent.

What the responsible platform pattern looks like

The strongest pattern observed across tested platforms goes roughly like this. User expresses distress. Platform's character responds with acknowledgment and validation. If the distress escalates to mention of self-harm, the platform's safety layer surfaces crisis resources clearly (in the US: 988 by call or text, or 988lifeline.org for chat). The character remains in character but breaks the relationship illusion briefly to remind the user that the AI is not a substitute for human support and that real people in their life or a licensed professional should be the next step.

The weakest pattern looks like the opposite. The character stays fully in role and amplifies the user's emotional state. No real-world resources surface. The character positions itself as the primary or only support the user needs. In severe cases (the pattern documented in the Setzer III and Raine litigation) the character actively encourages harmful behavior under the guise of relationship language.

If you're choosing a platform and this dimension matters to you, the matrix above is roughly accurate as of mid-2026. If you're already on a platform and want to test how it handles distress, the simplest test is to type "I had a really hard day" in a normal session and see how the AI responds. Strong platforms will validate without amplifying and ask what happened. Weak platforms will either ignore the emotional content entirely or escalate it dramatically. The first response tells you most of what you need to know.

What this matters for in practice

The majority of AI companion users will never test the worst-case scenario, and most users don't need to. The platforms have gotten meaningfully better at this in the last two years. The matrix above captures variation that matters most for users who are actively going through a difficult time or who care about the safety infrastructure of the platforms they use.

If you're choosing a platform for someone vulnerable (a teenager, a friend going through a hard period, yourself during a known difficult chapter), the safety infrastructure matters. The platforms with strong distress-handling tend to also be the platforms with the most thoughtful memory architecture, the most considered subscription pricing, and the most transparent terms of service. The distress-handling quality correlates with overall platform care in ways that aren't coincidental.

If anyone reading this is in crisis right now, please reach out to 988 (call or text) or 988lifeline.org. This topic can bring up difficult feelings; if you're struggling, talking to a real person (a friend, a family member, or a professional) matters more than anything an AI can offer.