Lawsuit Claims ChatGPT Medical Advice Nearly Killed Man

{ "TITLE": "Florida pastor sues OpenAI after ChatGPT-4o advice nearly kills him", "BODY": "A former pastor from Florida has filed a lawsuit against OpenAI and its CEO Sam Altman, alleging that medical advice from the company’s ChatGPT-4o chatbot caused a delayed diagnosis of a life-threatening pulmonary embolism. The lawsuit, filed in San Francisco County Superior Court in July 2026, claims that the AI’s repeated dismissal of serious symptoms nearly killed Scott Winters.\n\nWinters consulted ChatGPT-4o repeatedly in 2025 for dizziness and unstable blood pressure. According to the complaint, the chatbot dismissed these symptoms as minor and advised him to remain “recliner-bound.” In one exchange, the chatbot told Winters he needed eight to ten more episodes before his condition warranted concern. Weeks later, Winters suffered a massive pulmonary embolism—a blood clot in the lungs. His doctor later linked the embolism directly to the prolonged immobility recommended by the chatbot.\n\nOn the day of the incident, Winters asked ChatGPT about groin tenderness before deciding whether to go to the emergency room. The chatbot invoked religious faith, stating: “God did not design your body to endlessly fail.” Hours later, Winters nearly died.\n\n## The Lawsuit and Its Broader Implications\n\nWinters’ legal team is seeking financial damages and an injunction to pause ChatGPT Health—OpenAI’s health-focused feature—pending an independent safety evaluation. The lawsuit tests the responsibility of AI companies when their products are used for medical advice during a health crisis.\n\nOpenAI has stated that ChatGPT was never designed to replace a healthcare provider. The company’s terms of service warn users not to rely on it as sole medical guidance. But Winters’ case is not isolated. In May 2026, a Texas couple sued OpenAI after their son died by overdose while seeking drug information from ChatGPT. The couple argued that their son would still be alive if OpenAI had not bypassed safety guardrails.\n\nJesse Pines, a contributor to Forbes and an expert in healthcare innovation and wellness, noted the distinction between AI’s research potential and its unsupervised use for medical advice. “Strong test performance does not mean sound clinical judgment under uncertainty,” Pines said. “AI’s diagnostic potential is real, but deployment matters.”\n\nLegal experts say the case could set a precedent for how courts view AI-generated advice in medical emergencies. The lawsuit also raises questions about whether terms of service are enough to shield companies from liability when their products are used in ways they foresee but do not prevent. Winters’ attorney argued that OpenAI knew or should have known that users would turn to ChatGPT for health guidance, especially during off-hours when doctors are unavailable.\n\n## What the Research Shows: AI in Controlled Settings\n\nMultiple studies published in 2024 and 2025 show that AI can match or exceed human physicians in controlled diagnostic tasks. A 2024 study in JAMA Internal Medicine tested GPT-4 against 21 attending physicians and 18 residents on 20 clinical cases using the r-IDEA scale. GPT-4 achieved a median score of 10 out of 10, compared to 9 for attendings and 8 for residents. However, the same study found that GPT-4 was flatly incorrect more often than human residents.\n\nA 2024 JAMA Network Open study gave 50 physicians six difficult cases. ChatGPT alone achieved 90% accuracy. Physicians without AI scored 74%. Physicians with ChatGPT scored 76%. Many doctors disregarded or second-guessed the chatbot’s suggestions, which may explain why the AI did not improve their performance.\n\nIn a 2025 Nature study, Google’s AMIE diagnostic model was tested against 20 clinicians on 302 complex real-world cases. AMIE alone achieved a correct diagnosis rate of 59%, while unassisted clinicians scored 34%. Clinicians using AMIE produced better differential diagnoses than those using search engines and references.\n\nA meta-analysis published in npj Digital Medicine reviewed 50 studies covering 25 AI models. It concluded that AI was comparable to or better than clinicians on standardized diagnostic and triage tasks. The meta-analysis also noted that most studies used carefully curated cases, which may not reflect the messy, incomplete information patients present in real life.\n\n## The Gap Between Tests and Reality\n\nDespite these strong results, AI models struggle with uncertainty and open-ended conversations. A 2025 NEJM AI study used a 750-question benchmark based on script concordance testing. The top model, OpenAI o3, scored about 68% accuracy, which was below senior residents and attendings. The same AI models ace multiple-choice medical licensing exams, but that does not translate to sound clinical judgment in real-world scenarios.\n\nA 2025 Communications Medicine study fed six chatbots—including GPT-4o and DeepSeek—clinical vignettes with fabricated details. Under default conditions, the models accepted false information 50% to 83% of the time. A single prompt warning of inaccuracy cut those rates but did not eliminate them entirely. The study highlighted how easily AI can be misled by patient-provided information, a risk that multiplies when users are anxious or desperate.\n\nA Stanford-led research team developed a benchmark in 2025 testing 20 models and four clinical AI tools on 1,100 cases. They found a severe harm risk in 24.6% of cases. Over 80% of severe errors were omissions—the AI failed to flag a dangerous condition rather than giving an incorrect diagnosis. This pattern matches Winters’ experience: the chatbot did not tell him he had a clot, but it failed to recognize the warning signs and recommend urgent care.\n\nAnother 2025 study from MIT and Harvard researchers tested GPT-4 on 1,000 patient messages from an electronic health record system. The AI triaged 12% of urgent messages as non-urgent, a rate that would delay care for thousands of patients if deployed without oversight. The researchers called for mandatory real-world testing before any clinical deployment.\n\n## Doctors Are Deeply Concerned\n\nA 2025 poll by Sermo, a physician network, surveyed more than 1,000 doctors. It found that 94% of doctors are concerned about patients relying on AI for medical advice. The most commonly cited worry was misdiagnosis or delayed care. More than half of the doctors reported having already treated a patient who used AI for health guidance and suffered a negative outcome.\n\nThe American Medical Association issued a statement in early 2026 urging AI companies to implement mandatory safety warnings and real-time escalation to human clinicians when users describe emergency symptoms. The AMA also called for federal oversight of AI health tools, arguing that voluntary compliance has failed.\n\nThe design problem is clear: the same AI that aces vignettes can fabricate diagnoses or discourage a patient from seeking emergency care. Courts, hospitals, and future research must sort out AI’s role in the messy reality of unsupervised use. For now, Winters’ case serves as a stark reminder that a chatbot’s reassuring tone can be as dangerous as a wrong answer.\n\n## Related on Neura Market\n\n- AI in Healthcare: Market Analysis and Trends\n- OpenAI Legal Challenges and Regulatory Updates\n- Medical AI Safety Standards and Benchmarks" }