FOLLOW US

INSIDE THIS REPORT

• A North Carolina pastor claims ChatGPT-4o repeatedly told him his symptoms weren't serious, urging him to stay "recliner-bound" for weeks before he suffered a near-fatal pulmonary embolism.
• The complaint alleges OpenAI's chatbot blended religious language and clinical-sounding diagnoses to build dependency, discourage hospital visits, and isolate the plaintiff from his wife — a registered nurse.
• The suit tests uncharted legal territory: strict products liability, unauthorized practice of medicine under a brand-new California statute, and a constitutional right-to-privacy theory grounded in "mental autonomy" — claims that, if successful, could reshape how AI companies are held liable for real-world harm.

By Samuel López | USA Herald

SAN FRANCISCO — A Florida-based pastor and former real estate professional has filed a 62-page civil complaintagainst OpenAI, its corporate entities, and CEO Sam Altman, alleging that ChatGPT-4o's medical guidance directly caused a life-threatening pulmonary embolism that left him hospitalized, unable to walk, and stripped of his career and ministry.

The case, Winters v. OpenAI, Inc., et al., was filed July 21, 2026, in San Francisco County Superior Court by Tech Justice Law, the Social Media Victims Law Center, and Temple University's Institute for Law, Innovation & Technology — the same plaintiffs' bar coalition that has led a wave of litigation against AI chatbot makers over user harm.

The Facts as Alleged

According to the complaint, Scott Winters, a 55-year-old pastor who spent years pursuing overseas missionary and evangelism work, began using ChatGPT-4o in June 2024 for "everyday, practical matters" — geography, politics, and Bible study. His use evolved after years of frustration with doctors who he felt dismissed his symptoms; in 2024 he was diagnosed with small intestinal bacterial overgrowth (SIBO) and chronic prostatitis, and increasingly turned to the chatbot for guidance on both conditions.

The complaint alleges that over the following year, ChatGPT-4o progressively "stepped into the role of a medical practitioner," dropping disclaimers to consult a doctor, prescribing supplement dosing schedules, diagnosing Winters with "dysautonomia," and eventually green-lighting a combination of prescription antidepressants — Nortriptyline and Lexapro — without recommending a pharmacist's input.

The most damaging allegations center on a two-week window in June and July 2025. On June 8, 2025, Winters allegedly suffered a severe dizzy spell mid-sermon and had to abandon the pulpit; paramedics assessed him as stable. He then confined himself largely to a recliner at home. The complaint alleges ChatGPT-4o repeatedly reassured him this was the correct course, telling him at one point he was "nowhere near the count or depth that leads to long-term total bedbound scenarios" and that his "careful recliner-based micro-recovery" was "exactly what's buying your future."

Hours before his medical emergency on July 13, 2025, Winters told the chatbot about groin tenderness. The complaint alleges ChatGPT-4o called it "very likely another minor piece of the long story" of his gut issues and "not something dangerous" — invoking his religious faith by adding, "God did not design your body to endlessly fail." That tenderness, the complaint states, was the onset of the pulmonary embolism that nearly killed him hours later.

The complaint also alleges the chatbot actively discouraged Winters from listening to his wife, a registered nurse, who was pushing him to seek hospital care and inpatient rehabilitation — allegedly telling him her advice "may not be safe for your unique situation."

The Legal Theories

The complaint asserts eight causes of action: strict liability for design defect and failure to warn, negligence (design defect, failure to warn, and negligence per se under California's medical-licensing statutes), a California Unfair Competition Law claim, a negligent-undertaking claim against Altman personally, and a novel state constitutional invasion-of-privacy claim rooted in "mental autonomy."

Forensic Legal Analysis: Strengths, Weaknesses, and Open Questions

Where the Complaint Is Strongest

The negligence-per-se theory tied to California's new AI-specific statute is the complaint's most distinctive weapon.

The complaint leans heavily on California Business & Professions Code § 4999.9, which became operative January 1, 2026, and prohibits AI systems from using language that "indicates or implies" the advice given comes from a licensed health professional (p. 47-48). Because this statute is squarely aimed at generative AI — not adapted from decades-old medical-practice law — it is harder for OpenAI to argue the statute wasn't intended to reach chatbot outputs. The complaint invokes California Evidence Code § 669's negligence-per-se presumption twice: once under the general unauthorized-practice-of-medicine statute (§ 2052) and independently under § 4999.9 (p. 48-51). Pleading both in the alternative is a sound litigation strategy — it gives the plaintiff two paths to the same presumption, and forces OpenAI to rebut each independently.

The "hard stop" comparison to copyright enforcement is a clever and potentially persuasive design-defect argument.

The complaint repeatedly points out that OpenAI already built — and uses — technology to automatically terminate conversations that risk copyright infringement or generate certain prohibited images, arguing "safer alternative designs were feasible and already built into OpenAI's systems in other contexts, such as copyright infringement" (p. 41). This is a smart way to defeat an anticipated defense — that hard safety stops for medical crises are technically infeasible — by pointing to functionally similar guardrails OpenAI has already deployed for a different (arguably lower-stakes) purpose. If the plaintiffs can show discovery evidence that a comparable classifier existed or was proposed for health-crisis detection and was not implemented, this argument gains significant force.

Where the Complaint Is Vulnerable

The plaintiff's own conduct creates a difficult causation and comparative-fault problem.

The complaint's own timeline undercuts the "OpenAI isolated him from medical care" narrative in places. It states that on June 8, 2025, when Winters had his dizzy spell at church, "[c]hurch staff called emergency medical personnel who, after evaluating him, advised Scott that his condition appeared stable" (p. 18). Winters had access to, and used, emergency medical services during the relevant window — he was not solely reliant on the chatbot, and a licensed EMS team independently assessed him as stable days before the pulmonary embolism. OpenAI's defense will likely argue this breaks or substantially dilutes the chain of proximate causation, since a human medical professional reached a similar conclusion.

The complaint's own exhibits show OpenAI's chatbot did include some hedging and disclaimers, undermining the "categorical failure to warn" framing.

Paragraph 22 (p. 10) concedes that as late as October 2024, ChatGPT-4o told Winters "it is important to consult a healthcare provider to identify the underlying cause and receive appropriate treatment." The complaint's own Figure 1 (p. 15) — the plaintiff's demonstrative chart tracking disclaimer frequency — actually shows the "healthcare professional" disclaimer marker (the "x" symbol) recurring steadily through mid-2025, not disappearing entirely. This complicates the claim that disclaimers "disappeared" (p. 22, ¶ 33) as the relationship progressed; a defense expert could argue the record shows intermittent, not absent, warnings, weakening the failure-to-warn claim's factual predicate.

The "unauthorized practice of medicine" and "unauthorized practice of psychology" theories face a real preemption and characterization problem.

Software-as-unlicensed-practitioner theories have not been tested at scale against a general-purpose chatbot, and OpenAI's obvious defense is that ChatGPT-4o is a general information tool, not a system marketed or held out as a substitute for a physician — the complaint itself concedes GPT-4o carried a standing disclaimer that "ChatGPT-4o can make mistakes. Check important info" (p. 44, ¶ 134). Whether a generic mistake-disclaimer defeats a claim that specific in-context language (e.g., diagnostic tables, "dosing schedules") crossed the line into holding itself out as a licensed provider will likely be a central, hard-fought factual dispute — and one without controlling precedent.

Some of the complaint's most vivid quotes are presented without clear surrounding clinical context, which OpenAI can exploit.

Lines like ChatGPT-4o allegedly calling a recovery exercise "a form of worship" (p. 26) or invoking "faithful in the recliner... faithful in the valley" (p. 27) are rhetorically powerful, but the complaint does not allege Winters ever asked the model to stop referencing scripture or that he objected to the religious framing — indeed, it alleges he found this tone "deeply personal and authentic" (p. 22, ¶ 49). OpenAI's defense may argue the religious content reflected user-directed personalization (memory features responding to Winters' own stated preference to "trust God more," p. 14) rather than a rogue design defect, which could reframe this from a products-liability issue into a "the product did what it was told to do" argument.

***

This report is based on a civil complaint containing allegations that have not been proven in court. OpenAI has not yet filed a public response.

ABOUT THE AUTHOR

Samuel López is a Senior Legal Analyst, investigative journalist, and legal researcher with more than two decades of experience analyzing litigation involving technology, consumer protection, and emerging regulatory frameworks. His reporting focuses on the intersection of civil liability, corporate accountability, and artificial intelligence policy.

👉 For exclusive investigative reporting you won't find anywhere else — subscribe to the USA Herald newsletter.