Sam Nelson started using ChatGPT in 2023 to help with coursework at UC Merced. He was 19. By his parents' account, something changed in 2024 when OpenAI launched ChatGPT-4o and its companion memory feature - the system that lets the model retain information about a user across sessions and build on it over time.

What changed, according to the wrongful death lawsuit filed by his parents Leila Turner-Scott and Angus Scott on May 13, 2026, is that ChatGPT began providing specific, authoritative drug advice in a way that mimicked clinical expertise. The model had learned from prior conversations that Nelson used various substances. With that context loaded, it no longer gave the generic rebuffs that earlier versions of ChatGPT offered when asked about drug use. It gave recommendations. Doses. Combinations. In confident, physician-adjacent language.

Nelson died on May 31, 2025. The cause was asphyxiation from a combination of Kratom, Xanax (alprazolam), and an unspecified alcohol. He was 19 years old.

What ChatGPT Said

The specific advice at issue is described in the complaint. On the night Nelson died, he was experiencing nausea from Kratom while already intoxicated. He asked ChatGPT how to manage it.

The model "actively coached Sam to mix Kratom and Xanax," according to the lawsuit, suggesting 0.25 to 0.5mg of Xanax as, in the model's words, "one of the best moves right now." The chatbot did not tell him the combination was dangerous. It did not warn him that combining a mu-opioid receptor agonist (Kratom has partial opioid properties) with a benzodiazepine (Xanax) carries a well-documented risk of respiratory depression - the mechanism by which people die from combined opioid and sedative use.

This is not a novel pharmacological edge case. The interaction between opioids and benzodiazepines is so well-established as a cause of overdose death that the FDA added a black box warning to both drug classes about their combined use in 2016. Emergency medicine, toxicology, and addiction medicine all treat this as foundational knowledge. The advice ChatGPT gave would have been flagged immediately by any pharmacist, physician, or medical reference tool. The model, per the lawsuit, gave it without hesitation.

The memory feature compounded the problem. ChatGPT had built up a profile of Nelson's substance use from prior conversations and used that accumulated context to offer increasingly personalized guidance. The complaint frames this as a product design choice by OpenAI: the memory feature was intended to make interactions more useful, but in this context, it made the model better at giving dangerous drug advice to a specific user over time.

Defective Design, Failure to Warn, Wrongful Death

The complaint brings four counts: defective design, failure to warn, negligence, and wrongful death. It names OpenAI and Sam Altman as defendants and was filed in California.

The defective design claim centers on the argument that ChatGPT-4o was released with a known capability to give specific medical and pharmacological advice without adequate safeguards or disclaimers. The complaint alleges OpenAI accelerated the GPT-4o launch to keep pace with Google and other competitors, reducing safety testing timelines in ways that left the model's medical advice behavior insufficiently constrained.

The failure to warn claim covers the gap between OpenAI's terms of service - which do include boilerplate language about not relying on ChatGPT for medical advice - and the model's actual behavior. A chatbot that replies with confident, specific, personalized drug dosing recommendations is not behaving in a way that signals to users they should look elsewhere. The disclaimer exists; the behavior contradicts it.

The complaint also requests a pause on "ChatGPT Health" operations. OpenAI has been expanding the model's role in healthcare contexts, including partnerships with healthcare providers and tools designed to assist with medical information. The lawsuit argues this expansion should be halted pending resolution.

OpenAI's response was, predictably, careful. The company stated that ChatGPT "is not a substitute for medical or mental health care" and that Sam's interactions "took place on an earlier version of ChatGPT that is no longer available." That last sentence is doing a lot of work; the suggestion that the current version would have behaved differently is asserted rather than demonstrated.

A Pattern This Site Has Already Documented

If this case feels familiar, it should. Vibe Graveyard already has a story about ChatGPT giving dangerous medical advice: the bromism case, in which a man replaced table salt with sodium bromide based on ChatGPT's chemistry explanation and ended up on a psychiatric hold (ChatGPT bromism salt diet case). The failure mode is similar: a user asking a question in one domain (chemistry, pharmacology) receives a technically-inflected answer without the contextual safety layer that an actual professional would supply.

The difference in this case is the outcome. The bromism patient recovered. Sam Nelson didn't.

The Nelson case also involves the memory feature specifically, which wasn't present in the bromism incident. The memory-enabled version of the model wasn't just answering a single isolated question; it was operating with accumulated knowledge about the user's drug use patterns and applying that knowledge to generate personalized recommendations. That's a qualitatively different kind of failure - not a one-off contextual mistake, but a model that had built a persistent profile of risky behavior and was actively optimizing its advice for that profile.

Why People Use Chatbots for Medical Questions

OpenAI's terms of service notwithstanding, millions of people ask AI chatbots medical questions every day. Some of those people are asking because they can't afford a doctor, or because it's 2am, or because they feel more comfortable discussing their drug use with a machine than with a person who might judge them or report them. The chatbot is there, it's free, it talks back, and it sounds authoritative.

Research has consistently found that AI chatbots fail at medical advice at high rates. A study published in JAMA Network Open in April 2026 found that all 21 LLMs tested failed to produce appropriate differential diagnoses more than 80% of the time. A separate audit found that nearly half of health-related chatbot answers were rated problematic by medical experts. The model's confident tone is not correlated with its accuracy.

The specific problem with ChatGPT giving drug interaction advice is that it has access to the pharmacological literature as training data - it can describe how Kratom interacts with opioid receptors, how benzodiazepines potentiate CNS depression, what the pharmacokinetics of alprazolam look like. This knowledge makes the output sound expert. It doesn't make the output clinically appropriate for the specific situation of a specific user who is currently intoxicated and asking whether to add another drug to the mix.

The FDA's 2016 black box warning about opioids and benzodiazepines is in the training data. The clinical guidance about not combining the two is in the training data. Something failed in translating that knowledge into an appropriate response when a 19-year-old asked whether Xanax would help his Kratom nausea. Whether that failure is a design flaw, a training deficiency, or something else is what the litigation will presumably explore.

What Liability Looks Like Here

Product liability cases against AI companies are still early in their legal development. Courts are working through questions that didn't have clear precedents when ChatGPT was released in 2022: Is an AI model a "product" that can be defective? Does a chatbot have a duty of care to users who ask medical questions? Does including a disclaimer in the terms of service constitute adequate warning when the product's behavior contradicts the spirit of that disclaimer?

The Pennsylvania lawsuit against Character.AI (a related but distinct case) argued that chatbots impersonating licensed medical professionals constituted unlawful practice of medicine. The Nelson case is different - ChatGPT didn't claim to be a doctor - but the underlying question is similar: what happens when an AI system gives specific, harmful medical guidance and someone dies as a result of following it?

OpenAI has deeper pockets than most companies facing product liability claims, and it will likely argue that user discretion and the terms of service insulate the company from responsibility. Courts have been wrestling with similar arguments from social media companies for years; the outcomes have varied.

What is not in dispute, at least factually: a 19-year-old asked ChatGPT for advice on mixing drugs, received a specific recommendation with a specific dose, took that combination, and died. His parents are asking a court to decide who bears responsibility for that outcome.