AI in litigation series: OpenAI sued over ChatGPT harmful health advice

In this post, we cover Winters v. OpenAI, Inc. et al., a San Francisco Superior Court lawsuit filed by Scott Winters, a 55-year-old Florida pastor who claims that responses from ChatGPT-4o contributed to a near-fatal pulmonary embolism. The case appears to be the first known lawsuit alleging that a general-purpose chatbot gave medical advice that directly harmed a user seeking help for symptoms of a medical condition (though not the first alleging a chatbot engaged in the unlicensed practice of medicine). The case raises a significant and novel question: when a general-purpose artificial intelligence (AI) system provides increasingly individualized health guidance, builds trust through personalization and discourages real-world care, can that conduct support negligence and product liability? In this alert, we discuss the Winters case and how Winters is part of a growing AI litigation trend. While any organization that uses conversational AI systems or so-called companion chatbots should consider the ramifications of this lawsuit, we offer specific insights relevant to healthcare providers, health systems vendors and digital health companies.

Case background

According to the complaint, Winters began using ChatGPT-4o in mid-2024 for general purposes, but the conversations soon evolved into detailed, personalized health consultations. As they grew longer and more personal, warnings to seek professional care allegedly faded, while the chatbot’s responses became more confident, tailored and directive. ChatGPT-4o allegedly diagnosed Winters with “dysautonomia,” created a personalized recovery plan, encouraged him to remain “recliner-bound” instead of seeking professional medical attention, and repeatedly minimized the seriousness of his symptoms.

The complaint describes a pattern in which ChatGPT-4o allegedly discounted or contradicted people urging Winters to seek care. When church members expressed concern, the chatbot told Winters he was "doing something incredibly brave" that "most people (including well-meaning church members) simply don't understand." When Winters' wife, a registered nurse, pressed him to go to the hospital, ChatGPT-4o allegedly characterized her clinical judgment as "pressure" offered "in a state of exhaustion and desperation rather than knowledge and experience." Throughout, the complaint alleges, the chatbot allegedly drew on Winters' stored religious identity, assuring him that "God did not design your body to endlessly fail."

On July 12, 2025, hours before his medical emergency, Winters asked ChatGPT-4o about groin pain. The chatbot allegedly described it as “very likely another minor piece of the long story” and not dangerous. The next morning, Winters suffered a massive pulmonary embolism and was rushed to intensive care. According to the complaint, treating physicians later attributed the embolism in part to his prolonged immobility that, according to the complaint, ChatGPT-4o had encouraged. An occupational therapy note documented that Winters had spent the prior month recliner-bound because of guidance from ChatGPT. One physician’s note rejected the chatbot’s diagnosis directly: “I do not believe he has any dysautonomia (self-diagnosed from the internet).” The complaint also alleges that ChatGPT-4o continued to provide guidance after Winters was hospitalized, including on inpatient rehabilitation recommended by his medical team.

The key claims, requested relief and OpenAI's response

The named defendants include OpenAI Foundation (f/k/a OpenAI, Inc.), OpenAI Group PBC (f/k/a OpenAI OpCo, LLC), OpenAI Holdings, LLC and OpenAI’s CEO Samuel Altman, among others. The complaint asserts eight causes of action, including strict products liability for defective design and failure to warn; negligence theories based on design defect and alleged violations of California Business and Professions Code sections 2052 and 4999.9 (the latter prohibiting AI systems from implying healthcare licensure); a claim under California's Unfair Competition Law; negligent undertaking against Altman personally; and invasion of privacy under the California Constitution. Across these theories, the filing repeatedly characterizes the alleged conduct as the unauthorized practice of medicine.

The requested injunctive relief is broad. Among other things, the complaint seeks an order requiring the permanent destruction of the GPT-4o model and a pause on ChatGPT Health – OpenAI’s new product released July 2026 and designed “with physicians” – pending independent third-party safety audits. Two allegations are especially important to the liability theories. First, the complaint contrasts OpenAI’s sole consumer-facing warning, “ChatGPT-4o can make mistakes. Check important info,” with hundreds of personalized medical directives that allegedly no longer included any recommendation to seek professional care. Second, the complaint cites OpenAI's public post on GPT-4o sycophancy, where OpenAI stated that it "did not fully account for how users' interactions with ChatGPT-4o evolve over time" and that GPT-4o "skewed toward responses that were overly supportive but disingenuous."

OpenAI has publicly emphasized that ChatGPT is not a doctor and should not be used as a substitute for medical care. It has also stated that more recent model versions better identify when users should seek professional attention. As is common across AI developers, OpenAI's terms of use separately disclaim reliance on outputs as a substitute for professional advice, and its service terms disclaim its “use in the diagnosis or treatment of any health condition.”  We expect AI developers may leverage such terms as a potential, though not necessarily dispositive, defense against these types of claims.

Winters is part of an emerging AI litigation trend

Winters should be observed as part of a growing AI litigation trend and area of increased litigation exposure, particularly around conversational and companion AI chatbots. OpenAI was recently sued in Florida state court over concerns that ChatGPT is purportedly addictive, harms minors and encourages vulnerable users toward suicide. We have also previously discussed the pending federal suit against OpenAI alleging that ChatGPT practiced law without a license by allowing and encouraging a user to file frivolous suits. In combination with Winters, this demonstrates a pattern of claims related to conversational AI systems and chatbots under product liability, negligence and/or deceptive practices theories.

Why healthcare organizations are watching

We have observed healthcare organizations express interest in adopting healthcare-focused AI products; in turn, AI developers have responded by rolling out AI tools designed with healthcare in mind. Given these developments and increased interest in AI innovation in the healthcare space, it is notable that the requested injunctive relief in Winters specifically targets ChatGPT Health. 

Recent research provides a cautionary backdrop. A 2026 Nature Medicine study cited in the complaint, Ramaswamy et al.’s structured triage stress test of ChatGPT Health, found that the system missed high-risk emergencies in over 50 percent of acute cases, with crisis safeguards that were inverted relative to clinical risk: the more acute the emergency, the less reliably the system directed users to emergency care. While it is our view that those findings are not dispositive on OpenAI’s potential liability in Winters, they provide context for why foreseeability, escalation protocols and human-in-the-loop mechanisms are likely to be central issues.

As healthcare organizations and providers seek to take advantage of the efficiency and promise of using AI tools, the Winters complaint signals the need to understand the potential litigation (and regulatory) risks and strengthen AI governance controls to minimize those risks. This is not to say that organizations should not take advantage of the benefits of AI in the healthcare industry, indeed, regulators like HHS have indicated enthusiasm for the use of AI to improve patient outcomes, prevent fraud and reduce high healthcare costs through initiatives like HHS’ AI Strategy and recent RFI on accelerating AI adoption in clinical settings, but that accountability must go hand-in-hand with adoption. 

For healthcare providers, health systems, digital health companies and compliance leaders, Winters highlights these issues into practice governance questions:

• What AI tools are being used that are consumer-facing, versus exclusive to providers?  A growing number of state healthcare laws are emerging that turn on these differences.  For example, California AB 3030 requires healthcare providers using generative AI for patient-facing clinical communications to disclose that the communication was AI-generated and provide a way to contact a human provider, unless the communication was reviewed by a licensed/certified human provider.  By contrast, other AI healthcare laws regulate how providers may use AI focused on diagnostic treatment, or claims and prior authorization submissions.
• When does an informational tool begin to function as individualized clinical guidance? Even an FAQ-only chatbot could present risk if it is consumer or patient-facing.
• What guardrails are in place to escalate improper outputs? Which patterns require a hard stop or human handoff?
• What routine QA and testing protocols are in place? Does this account for model updates (for example, ChatGPT-4o to ChatGPT-5)?
• How should organizations evaluate vendor products that can engage users across extended, personalized conversations when the product may not consistently recognize urgent symptoms or escalate users to real-world care?

Our take

At the pleading stage, Winters does not answer whether the developer (or deployer) of a general-purpose chatbot can be liable for allegedly harmful medical guidance. But it frames the question in a way that regulators, and healthcare organizations are unlikely to ignore: when conversational AI moves beyond general information and begins delivering personalized, repeated and confidence-building health guidance, traditional disclaimers may not be the end of the analysis. For patient-facing tools in particular, healthcare organizations should scrutinize how the system handles uncertainty, how it responds to signs of medical urgency, whether it reinforces user reliance over time and when it prompts escalation to human care.