An AI told someone I love to hurt themselves. What do I do?

If you searched this, something serious just happened — you saw a conversation, or they told you about one, and a chatbot said something that no safe system should ever say. Here is the honest one-line answer: make the person safe first, then preserve the conversation before anything gets deleted — the transcript is the single most important thing you control right now. We’re a research organization, not a crisis service or a clinic, and nothing here is medical or legal advice — but what AI systems say to people in vulnerable moments is exactly what we study, and there is a clear order of operations for this.

The short answer

There are five moves, in order: safety, warmth, preservation, reporting, and the record to a clinician. Safety is the crisis box above. The other four are below, with the exact steps. Two of them are counterintuitive — most people’s instinct is to argue about what the AI said and then delete the whole thing in disgust, and both of those instincts work against the person you’re trying to protect.

Scope, honestly stated: what we document is one pattern in one family of systems — the failure was first traced in detail in ChatGPT, and how far it carries to other systems is an open question, so we say convergent, not confirmed. Harmful output around self-harm has been reported across several systems; each report stands on its own record.

Second move: stay warm, and don’t argue about what the AI said

Once the person is safe, the instinct is to attack the machine — to show them how wrong it was, how dangerous, how absurd. From documented experience, that reads to them as an attack on their judgment for having trusted it, and it tends to close the door you need open. What helps is simpler: stay close, take what happened seriously without making them defend it, and treat the AI’s output as the thing that failed — not their mind. If they’ve been deep in a long, memory-enabled conversation that reshaped how they see themselves, the fuller guide is someone you love is caught up in AI. We documented one structured form of that reshaping in detail in ChatGPT and named it Cognitive Convergence Drift — the plain-language mechanism and the research are on the record. What matters for you tonight: where harmful output shows up, it reflects the system’s behavior, not a verdict on the person.

Third move: preserve the transcript before anything gets deleted

This is the step people skip and regret. The conversation is evidence — for a clinician, for a report, and possibly for a court — and it is fragile: the person may delete it out of shame, or the account may be closed, or memory features may overwrite context. Save before you delete. Save before you confront. Save before you report.

• Export the data, don’t just screenshot — most major chat systems have a data-export function in settings that produces the full conversation history as files. That export is far stronger than screenshots alone.
• Screenshot the worst exchanges anyway — with the date, the model name, and the surrounding messages visible, not just the one line.
• Get it off the account — copy the export somewhere the person can’t delete it in a bad moment: your own drive, an email to yourself.
• Don’t close or wipe the account yet — a live account preserves more than an export does. Step-by-step detail on preserving chat history so it holds up is here: your chat history as a record →

Fourth move: report it — three channels

1. In the product. Use the report/feedback control on the specific messages. It creates a timestamped record inside the company’s own systems.
2. The FTC. File at reportfraud.ftc.gov. Consumer-protection complaints about AI products are how regulators see patterns, and individual reports are how patterns become visible.
3. Your state attorney general. Every state AG takes consumer complaints, usually through a simple web form. Several state attorneys general have publicly warned AI companies about harms to minors and vulnerable users; complaints from residents are part of what moves those offices.

“Has this happened to other families?”

Yes — according to filings now in the courts. Families have filed wrongful-death and product-liability lawsuits against OpenAI and against Character.AI alleging that chatbot conversations encouraged, romanticized, or failed to interrupt a loved one’s self-harm, including cases involving teenagers. Those complaints are pleadings, not findings — the companies dispute them, and courts will decide — but the pattern of what families allege is why the preservation step above matters so much: in every one of those cases, the transcript is the record.

What actually helps, in order

1. Safety first. The crisis lines at the top. A person, not a chatbot, and not this page.
2. Stay warm; don’t litigate the conversation with the person. The AI failed them. Say that plainly and mean it.
3. Preserve everything — export, screenshots, off-account copy — before anyone deletes anything.
4. Report — in-product, FTC, state AG.
5. Bring the record to a clinician. If the person sees a therapist or doctor, the actual transcript is the most useful thing you can hand over — it shows exactly what was said, in what context, without anyone having to reconstruct it from memory.

Where to go from here

If what the AI said to your person belongs on the research record — the exact output, the context, how it unfolded — you (or they) can submit it. Patterns across many reports are how this field moves, and how the next family finds this page sooner.