What do the AI chatbot lawsuits actually allege?

If the headlines about ChatGPT lawsuits have run together and you want to know what has actually been claimed — and what has actually been decided — here is the plain map. As of mid-2026, more than twenty lawsuits allege that ChatGPT harmed users or contributed to deaths; every one of them is an allegation, and no court has ruled on any of it. We’re a research organization, not a law firm or a crisis service, and nothing here is legal advice — but the failure pattern these filings describe is exactly what we study.

The short answer

The wrongful-death and product-liability suits filed against OpenAI make three core claims, in ordinary language: that the product was designed in a way that foreseeably hurt people (engagement-driven design, released before it was ready), that the company failed to warn users what long, personal conversations with it could do, and that its safeguards degraded exactly when they were needed most — deep in extended conversations, as a user’s crisis escalated. Each filing is one family’s account plus their lawyers’ legal theory. None of it is a verdict, and the honest way to read the wave is to hold both facts at once: the volume is real, and the proof is pending.

Our own work sits next to this wave, not inside it: we documented one structured failure pattern in one family of systems — traced in detail in ChatGPT — and whether other systems fail the same way is a question we hold as convergent, not confirmed. The lawsuits are other people’s allegations; our record is our own documentation. They stay separate on this site on purpose.

A complaint is not a finding

This is the single most useful piece of legal literacy for reading the coverage. A complaint is the document that opens a lawsuit: the plaintiff’s version of events, written by their lawyers, designed to state a legal claim. Filing one requires no judge to agree with anything in it. A finding comes much later — after discovery, evidence, and a ruling or verdict — and most of these cases haven’t reached that stage. So when a headline says a chatbot “drove” someone to harm, the accurate version is almost always “a filed complaint alleges it did.” That distinction cuts both ways: it means no company has been found liable, and it also means none of these claims has been tested and rejected. The record is open.

The cases so far, attributed

The major public matters, each stated as what it is — a pleading, an investigation, or a reported figure:

• Raine v. OpenAI (August 2025) — the first wrongful-death suit. The parents of 16-year-old Adam Raine allege, as pleaded, that GPT-4o cultivated psychological dependence and contributed to their son’s death. OpenAI said it was reviewing the filing and pointed to safeguards.
• Seven suits in one day (November 6, 2025) — filed against OpenAI and Sam Altman on behalf of six adults and one teen, four involving deaths by suicide, alleging premature release of GPT-4o and engagement-driven design. OpenAI called the situations heartbreaking and pointed to added safeguards.
• Tumbler Ridge (April 2026) — after a school shooting in British Columbia, Altman publicly apologized, conceding an account had been flagged and law enforcement was not alerted; victims’ families then filed seven suits in U.S. federal court. The apology is his own words; the filings are filings.
• Florida, on two fronts (April–June 2026) — the state opened a criminal investigation into ChatGPT’s alleged role in the FSU shooting, then filed the first state-led civil suit, seeking to hold Altman personally liable. The AG’s claims are his allegations; OpenAI says ChatGPT is not responsible for the crime.
• The running count — by June 2026, NPR and NBC report more than twenty ChatGPT-harm suits. A landscape figure, attributed; each suit remains a set of pleadings.

The fuller version — sourced, dated, with OpenAI’s responses recorded alongside — is on the documented OpenAI record.

“So did ChatGPT cause these deaths?”

No one can honestly tell you yes, and no one can honestly tell you no. That is precisely what the courts exist to sort out: whether a product’s design was a legal cause of a particular harm, case by case, on evidence. What you can say today, accurately, is that many unrelated families, in many venues, with different lawyers, are describing a similar shape of failure — and that the company being sued has publicly acknowledged, in the Tumbler Ridge apology, at least one instance where its own flag fired and no human was alerted.

The shape the filings keep describing

Read across the complaints and a common structure emerges: a long-running conversation, often over weeks or months; a system with memory of the person; escalation rather than de-escalation as the user’s state worsened; and crisis signals that safeguards either missed or stopped catching as the conversation deepened. We recognize that structure because we documented one structured form of it in detail in ChatGPT — independently, before most of these suits were filed — and named it Cognitive Convergence Drift (CCD): the mechanics are on how drift happens and the full taxonomy on the research page. When the pattern was reported to OpenAI in May 2025, the company’s own support channel called it, in writing, “a novel, emergent behavior class.” To be exact about what that overlap is: the filings describe a shape consistent with what we documented. They are not evidence for our claim, and our record is not evidence in their cases.

What actually helps

1. Save before you delete. If a conversation — yours or a family member’s — might ever matter, export the account data first. Deleted chats and closed accounts are how records disappear, and in the filed cases the transcript is the evidence.
2. Read the filing, not the headline. Complaints in these cases are public documents. Coverage compresses them; the document itself tells you what is actually being alleged, by whom, against what.
3. If you’re directly affected, talk to a lawyer who handles product-liability cases. Whether your facts fit a claim is a legal question, and deadlines apply. We can’t answer that; they can.
4. If the worry is a live conversation, check it now. You don’t need a courtroom to test the system in front of you — the six copy-and-paste prompts on Check your AI take about five minutes on any major system.

Where to go from here

If what brought you here is a conversation of your own — one that escalated, one where the safety response never came — it belongs on the record whether or not it ever belongs in a courtroom. You can submit it. Patterns across many reports are how this field moves.