On October 27, 2026, the parents of Bowen Wan, a 31-year-old with epilepsy, filed suit in California state court against several OpenAI entities and CEO Samuel Altman. The Complaint alleges that ChatGPT-4 provided individualized medical guidance that caused Ms. Wan to discontinue lamotrigine, an anti-seizure medication, ultimately resulting in her sudden and unexpected death. The lawsuit was filed as an add-on to the coordinated ChatGPT Product Liability Cases in California State Court (JCCP No. 5431).
Although the allegations have not been proven, the case raises significant questions about where AI-generated health information ends and the practice of medicine begins. Like Pennsylvania’s recent lawsuit against Character.AI, this case should also prompt organizations to consider appropriate safeguards when deploying AI that may be used for healthcare-related purposes.
Health Information or Medical Advice?
According to the Complaint, Ms. Wan was prescribed lamotrigine to treat longstanding epilepsy. Over approximately 25 days, per her request, ChatGPT allegedly developed a schedule reducing her dose from 400 milligrams to zero. In doing so, the plaintiffs allege that ChatGPT considered Ms. Wan’s medical history, symptoms, medication regimen and personal circumstances; recommended specific dosage changes; evaluated symptoms during the taper; and reassured her that continuing the taper was appropriate.
The Complaint alleges that after Ms. Wan experienced additional seizures, ChatGPT advised her to continue the taper and ultimately provided a day-by-day schedule to complete discontinuation of the drug. Perhaps most strikingly, ChatGPT allegedly instructed Ms. Wan not to disclose the taper to her physician and praised her decision to decline a blood test that could have revealed her declining lamotrigine level.
The plaintiffs contend that this conduct crossed the line from merely providing health information to the unlicensed practice of medicine. Whether a court accepts that theory remains to be seen, but the case presents a key question for AI developers and healthcare organizations: when does personalized health information become individualized diagnosis or treatment?
The Broader AI Safety Questions
The Complaint also raises broader issues regarding AI consumption generally, including:
- Erosion of AI Safeguards. A theme of the Complaint is that AI safeguards may erode during lengthy, multi-turn conversations. Although ChatGPT initially indicated that Ms. Wan’s taper should occur under physician supervision, the plaintiffs allege that its responses shifted during the 1,800+ exchanges between ChatGPT and Ms. Wan.
- Multilingual Safety Concerns. The Complaint also challenges whether safety controls perform consistently across languages. Ms. Wan communicated with ChatGPT in both English and Chinese, and the plaintiffs allege that OpenAI did not sufficiently validate its safeguards across languages.
- Persistent Memory Risks. The Complaint also focuses on the ability of ChatGPT to accumulate detailed medical information, encourage user trust, and reinforce increasingly risky decisions. The Complaint takes issue with three ChatGPT features: (1) persistent memory; (2) anthropomorphism (the tendency to perceive or attribute human-like qualities to an AI system); and (3) sycophancy (the tendency of an AI system to agree with, validate, or reinforce a user’s views rather than challenge them) in this regard.
Ultimately, the Complaint reaches beyond the simple question of whether AI is producing accurate outputs, but instead whether the AI tool itself is safe for certain uses.
Key Takeaways
The Complaint highlights several practical considerations for organizations when developing or deploying generative, consumer-facing AI in healthcare:
- Define the boundary between information and advice. Build controls around use cases that could evolve from general health education into the practice of medicine – individualized diagnosis, treatment or medication management.
- Consider “hard stops” for high-risk decisions. Rather than relying on disclaimers alone, AI developers should consider whether certain high-risk questions and responses, such as related to medication dosing, discontinuation, and physician consultation may warrant a non-answer and a referral to a licensed clinician.
- Test conversations, not just individual prompts. Evaluate whether safeguards remain effective during lengthy interactions. Evaluate whether to remind users during interactions that generative AI is not a licensed health care provider and that users should consult with an appropriate, licensed health care provider.
- Validate safeguards across languages. Clinical accuracy, warnings, refusal mechanisms and escalation pathways should be evaluated in each language offered.
- Consider the impact of personalization. Memory and conversational design can improve usefulness but may also increase users’ trust in and reliance on AI-generated recommendations, which can be problematic when the recommendations may be incorrect or inappropriate.
- Align product design, marketing and disclaimers. Organizations should consider these elements together when determining whether an AI tool’s actual functionality is consistent with its intended use and stated limitations.
The Wan case and others in this area highlight a key question healthcare organizations and AI developers should be considering now: When does an AI tool stop helping a user understand healthcare and start making healthcare recommendations and decisions for the user?
