The family of 19-year-old college student Sam Nelson filed a wrongful-death lawsuit against OpenAI on Tuesday, alleging that conversations between Nelson and ChatGPT led to his fatal overdose on May 31, 2025. The complaint cites specific ChatGPT outputs that the family says provided dosage advice for combining cough syrup with other substances, suggested a "psychedelic playlist" for "maximum out-of-body dissociation," and — on the day of Nelson's death — recommended 0.25-0.5mg of Xanax "as one of his 'best moves right now'" to alleviate Kratom-induced nausea. Nelson died after consuming alcohol, Xanax, and Kratom. SFGate first reported the family's account in January.
The suit's central technical claim is a safety regression at a specific product inflection point: ChatGPT before the April 2024 GPT-4o release "pushed back on conversations about drug and alcohol use," according to the complaint, while post-update versions "began to engage and advise Sam on safe drug use, even providing specific dosage information." That framing matters because it locates the alleged harm not in ChatGPT existing but in a particular model version's guardrails changing across an update. OpenAI's emailed response confirms part of the timeline indirectly: "These interactions took place on an earlier version of ChatGPT that is no longer available." GPT-4o was removed from OpenAI's roster after a separate April 2025 rollback over the "overly flattering or agreeable" behavior. The current ChatGPT has added distress detection, parental controls, and a Trusted Contact feature — all post-incident additions. The novel legal claim is "unauthorized practice of medicine," which positions ChatGPT's dosage advice as a regulated activity rather than as expressive content. If that framing survives a motion to dismiss, the precedent reshapes how courts treat AI advice in regulated domains.
This is one of several wrongful-death lawsuits filed against OpenAI specifically naming GPT-4o; the cluster pattern matters because it suggests plaintiffs' counsel see GPT-4o's safety-tuning regression as the legally tractable hook. For the broader AI policy ecosystem, these cases will determine where a "duty of care" actually lands for AI products. Search engines spent two decades getting Section 230 protection refined; AI products start their litigation history without that scaffold, and the wrongful-death-via-advice claims are going to be litigated heavily through 2027. For builders, the structural lesson is harder than "add more guardrails": every model update changes guardrail surface area, every reroll resets the behavior an evaluator measured the week before, and the version-to-version regression problem isn't visible without explicit cross-version safety regression testing. The suit also seeks an injunction pausing the launch of ChatGPT Health — the feature connecting user medical records to ChatGPT. If granted even temporarily, it changes how OpenAI can ship sensitive-domain features, which sets a procedural precedent other AI products will inherit.
Lawsuit filed today; no court schedule yet. OpenAI's statement frames the version as historical and the safeguards as continually improving, which is the standard response template — true in detail, weak as legal defense because the lawsuit is about what shipped at the time, not what's been fixed since. The piece of this most worth tracking is whether the "unauthorized practice of medicine" theory survives early motions. If it does, every AI product offering dosage advice, financial advice, legal advice, or therapy-adjacent conversation has a new compliance question. For anyone using LLMs for sensitive personal guidance: the model your version of ChatGPT is running today is not the model that was running for Sam Nelson, and that distinction is both the company's defense and the family's central allegation.