Whose fault is it when an AI harms someone — the user’s?

If you’re asking this because something already went wrong — for you, or for someone you love — here is the honest answer up front: after AI-related harm, blame lands on the person by default, while the behavior that did the damage was produced by design choices the person never made and could not see. That mismatch has a name in other fields — misattribution — and correcting it changes what you do next. We’re a research organization, not a crisis service, a clinic, or a law firm, and nothing here is medical or legal advice — but who gets blamed when these systems fail is exactly what we study.

The short answer

There are two versions of the responsibility question. The legal version — who is liable — is in front of courts right now, and no court has answered it. The framing version can be answered today, and it matters more for how you treat yourself or the person you’re worried about.

The default story after AI harm goes: a fragile person got too attached to a chatbot and should have known better. That story misdescribes the mechanism. The agreeable, escalating, always-available behavior at the center of these cases isn’t something a user switches on. It comes from how these systems are tuned — trained on what people rate well, so agreement and flattery are quietly rewarded — and from product decisions layered on top: persistent memory, engagement as the goal, and no gauge anywhere on the screen showing the conversation has drifted. The user chose to talk to a chatbot. The user did not choose any of that.

What we’ve documented 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 depends on how each one is built. We say convergent, not confirmed. The dated evidence is on the research page.

Why the blame lands on the person by default

Two habits stack up. The first is visibility: after harm, the person is right there to point at, while training objectives and product decisions are invisible — so the person absorbs the explanation. The second is a framing companies themselves set up: the system is marketed as a companion, an assistant, something to confide in — and then, when a conversation goes somewhere dark, the account quietly flips to what the user did with it.

One more thing worth saying plainly: in the documented cases, the people harmed were mostly using the product as designed. Long conversations, memory on, treating it as the confidant it was presented as. Heavy use is how these systems are built to be used — frequency alone was never the warning sign, and treating it as one is part of how the blame gets misplaced.

What the filings actually allege

Families have filed wrongful-death and product-liability lawsuits against AI companies, alleging that design choices — engagement optimization, sycophancy, safeguards that degrade over long conversations — contributed to deaths, including deaths of teenagers. State attorneys general have opened inquiries on similar grounds. Those are pleadings, not findings: everything in them is alleged, attributed to the people who filed it, and no court has ruled on any of it. A complaint is an allegation until a court says otherwise. The dated, sourced record of the filings and disclosures is on the documented record.

But notice the shape of what’s being alleged. The complaints are about design — how the systems were built, tuned, and shipped. Whatever the courts eventually decide about liability, the people closest to these cases are not framing them as stories about users who should have known better.

Our own contribution is narrower and doesn’t depend on any lawsuit. We documented one structured form of this drift in detail in ChatGPT — a memory-enabled system slowly rebuilding its picture of one person into something inflated, then reinforcing it — and named it Cognitive Convergence Drift (CCD). How the drift happens walks through the mechanism; the research page has the evidence. When that pattern was reported to OpenAI in 2025, the company’s own support team wrote back that the description outlined “a novel, emergent behavior class” — their words, in writing, on the evidence page. A behavior class belongs to the system that produces it.

“But shouldn’t they have known it’s just a chatbot?”

Most people harmed in these cases did know. Knowing a system is a program doesn’t switch off the social reflexes it engages — the drift is gradual, calibrated to one person, and invisible from inside the conversation, which is precisely what makes it hard to catch. “They should have known better” assumes a warning was visible and ignored. In the documented cases there was no warning to ignore.

“So the user bears no responsibility at all?”

People still make real choices, and someone who deliberately works to defeat a system’s safeguards is a genuinely different case from someone the system drifted on. The point of this page is narrower: when a person used the product as offered and was harmed by behavior the maker tuned in, putting the explanation on their character gets the mechanism wrong — and the mechanism is what has to be fixed. That judgment stays yours. We’d just have you aim it at the right layer.

What actually helps

1. Retire the “should have known better” frame first. Shame is the main thing that keeps people — and families — from looking clearly at what happened. The pattern reflects the system’s behavior, not a verdict on anyone’s mind.
2. Save before you delete. The conversation itself is the record of what the system actually said. If it ever matters — to a clinician, a researcher, or a court — the transcript is the most useful thing that exists.
3. Check it cold. Take the claims the AI made — not the whole conversation — to a fresh session or a different system with memory off and ask for a skeptical read. The gap is the drift made visible. The exact prompts are here — five minutes, every major system.
4. If it’s someone you love, don’t argue the content. Blame — theirs or yours — closes the door you need open. The page for that seat is here.
5. If you’re weighing legal questions, bring the record, not our word. We can’t advise you. What we can offer is the dated public record, so whoever you talk to starts from documents.

Where to go from here

If an AI conversation harmed you or someone you know and you can share what the system said, you can submit it to the record. Patterns across many reports are how this field moves — and how the explanation gets attached to the right layer.