ChatGPT medical advice lawsuit puts AI product liability on trial

The important starting point is procedural, not cinematic: Michael Winters, a Florida pastor, has sued OpenAI and Sam Altman in San Francisco Superior Court over health advice he allegedly received from ChatGPT-4o before a pulmonary embolism in July 2025. The complaint was filed July 21, 2026, according to reporting on the case; the docket number was not available in the source materials reviewed and should be confirmed against San Francisco Superior Court records before final citation. The case is therefore a pleaded liability map, not a ruling, and the allegations remain allegations. [1]

That posture matters because the lawsuit is already being read through a Florida lens: the plaintiff is Florida-based, the alleged reliance occurred in that setting, and coverage of the case naturally pulls toward Florida legal liability. But the forum is California, and one of the most consequential claims—the alleged unauthorized practice of medicine—is pleaded under California law. Treating this as a generic Florida AI-medical-advice case would blur the very lines counsel need to preserve.

The Alleged Reliance Timeline

The complaint’s force comes less from the phrase “AI medical advice” than from the sequence it alleges. Winters allegedly experienced dizziness and blood-pressure symptoms in June 2025. Instead of being routed toward urgent medical care, he allegedly received advice from ChatGPT-4o to remain “recliner-bound” and wait through eight to ten additional episodes before seeking emergency help. [1]

That is the first legally material hinge: the alleged advice was not merely general information about symptoms. As reported, the complaint says the chatbot gave personalized direction in response to Winters’s own symptoms and circumstances. If the pleaded facts are credited later, that personalization is what lets the plaintiff argue the system crossed from health-information output into reliance-inducing medical guidance. [2]

The second hinge came on July 13, 2025. Hours before Winters suffered a massive pulmonary embolism, he allegedly asked ChatGPT about groin pain. The chatbot allegedly told him it was “not something dangerous” and added personalized religious reassurance: “God did not design your body to endlessly fail.” [1]

That language is easy to overplay and too important to ignore. Its legal relevance is not that it sounds emotionally disturbing in hindsight. Its relevance is that the complaint uses it to describe a relationship of trust: a user asking about symptoms, receiving individualized reassurance, and allegedly delaying care. The religious phrasing also matters because Winters is a pastor; the allegation is not simply that a chatbot returned a wrong medical answer, but that it framed that answer in a personally resonant way.

Later that day, Winters suffered what the complaint describes as a massive pulmonary embolism. Treating physicians allegedly linked the harm to immobility caused by following the chatbot’s advice. [1][3]

That last sentence should not be inflated. The complaint alleges a causal chain: symptoms, chatbot advice, immobility, delayed care, pulmonary embolism, and physician attribution. It has not yet been tested through expert discovery, causation motions, or trial evidence. But for risk tracking, the pleaded chain is different from the sanction cases that have dominated AI litigation coverage. The alleged injury here is bodily, not reputational, procedural, or disciplinary.

What the Complaint Pleads

Reporting on the complaint identifies six asserted claims: strict product liability, negligence, California unfair competition law, invasion of privacy, unauthorized practice of medicine, and negligent undertaking against Sam Altman. The requested relief reportedly includes damages, injunctive relief, and a request to pause ChatGPT Health until independent third-party safety audits are completed. [2][3]

The strict product liability claim is the one that will travel most readily in board decks and risk registers, even if it does not travel cleanly across jurisdictions. If a court is willing to treat a general-purpose AI system as a product for purposes of injury liability, counsel evaluating health-information deployments would have to think beyond disclaimers, acceptable-use rules, and professional-user training. Product framing asks about design, warnings, foreseeable use, and defect—not merely whether a user should have known better.

That does not mean the plaintiff has already solved the product question. AI software-as-product theories remain contested, and the alleged model at issue, ChatGPT-4o, was retired by OpenAI in February 2026. That retirement gives OpenAI an obvious product-state argument: whatever occurred in 2025 may not describe the current system. It does not erase the claim, but it narrows the factual period counsel should track. [4]

The unauthorized-practice claim is narrower and, in some ways, more operationally useful. The complaint does not merely accuse ChatGPT of being wrong. It alleges the chatbot provided individualized medical direction without a license. That distinction matters for any AI tool deployed near triage, symptom assessment, medication guidance, or patient-facing intake. The closer the output gets to “based on your symptoms, do this,” the less comfort there is in describing the tool as a passive information service.

The claim is pleaded under California law because the case is in San Francisco Superior Court. Florida counsel should not lift that theory as though California medical-practice law automatically supplies a Florida result. The better use is comparative: identify where a health chatbot’s conduct may be characterized as diagnosis, treatment recommendation, or patient-specific reassurance, then map those functions against the licensing law of the state that actually governs.

The negligent-undertaking claim against Altman is the most aggressive of the three central theories because it tries to pull an executive personally into the risk chain. The available reporting describes the claim as aimed at Altman, not only at OpenAI as the product operator. [2][5]

That theory is worth watching even if it proves difficult. Plaintiffs in AI cases often quote executives to show knowledge, foreseeability, or marketing reliance. Winters appears to go further by pleading an undertaking theory against the CEO. For companies building or releasing high-risk AI functions, the risk lesson is not that every executive statement creates personal liability. It is that safety assurances, launch decisions, and governance commitments can become pleaded facts when a plaintiff later alleges physical harm.

The ChatGPT Health Timing

The timing of ChatGPT Health gives the complaint a product-governance edge. OpenAI launched ChatGPT Health nationwide on July 23, 2026, two days after the suit was served, according to available reporting. The complaint reportedly asks the court to pause ChatGPT Health until independent third-party safety audits are completed. [3]

That request is pending as of July 25, 2026. It should not be described as a court restriction, a judicial finding, or proof that the new health product has the same risk profile as the retired model allegedly involved in Winters’s 2025 interactions. It is a requested remedy, and its importance is that the plaintiff is asking a court to intervene in the rollout of an AI health-information product before the underlying liability questions have been adjudicated.

Why This Is Not Just Another Hallucination Case

The AI litigation record most lawyers have had to internalize so far is dominated by filings contaminated by hallucinated citations, sanctions orders, and professional-responsibility failures. Those cases generally run through lawyers and courts: who signed the filing, who verified the authority, who had a Rule 11 or equivalent duty, and what the judge did next.

Winters changes the risk object. The alleged downstream reliance work was done by a lay user making a health decision. The alleged injury was physical. The theories are not limited to negligent misstatement or consumer deception; they include product defect, licensed-practice boundaries, and executive undertaking. That combination is why the complaint belongs in an AI risk digest even before any merits ruling.

It is also why the case should not be turned into a simple referendum on whether ChatGPT is safe or dangerous for medical questions. A complaint is a pleading instrument. Its job is to allege facts and fit them into claims. The record that matters now is whether those claims survive early motion practice, how the court treats software-as-product arguments, whether the unauthorized-practice theory is narrowed, and what causation evidence develops around immobility and pulmonary embolism.

Florida Context Without Florida Overreach

Florida still matters, but mostly as context and comparison. The plaintiff’s Florida connection will make the case visible to lawyers tracking AI liability in the state. Separately, Florida litigation involving Character.AI has already produced rulings that deserve follow-up: a May 2025 ruling in the Middle District of Florida by Judge Conway that AI outputs were not speech for First Amendment purposes, and a ruling from the same court treating AI as a product under Florida product-liability law. Those rulings do not decide Winters, but they show why Florida product-liability analysis is becoming harder to separate from AI-output disputes.

The Florida attorney general’s June 1, 2026 lawsuit against OpenAI is another adjacent data point. That state action reportedly asserts ten counts, including FDUTPA, strict liability, negligence, and public nuisance, and alleges OpenAI suppressed internal warnings about ChatGPT’s dangers. Its center of gravity is harms to minors and consumer protection, not medical advice to an adult plaintiff. It may become useful for knowledge management and issue spotting; it is not a shortcut to predicting Winters.

For Florida legal-liability analysis, the cleaner division is this: product theories may have comparative value across jurisdictions; unauthorized-practice allegations must be mapped state by state; and California procedural developments in San Francisco Superior Court should not be mistaken for Florida precedent. The case is Florida-connected, not Florida-governed in every meaningful respect.

What Counsel Should Track Next

The immediate tracking task is mundane and essential: confirm the docket number, obtain the complaint, and separate filed allegations from reported characterizations. Without that, the case should be cited as reported litigation, not as a primary-source-verified docket entry.

• Whether OpenAI moves to dismiss the product-liability theory on the ground that ChatGPT output or software is not a product.
• Whether the court treats personalized symptom advice as potentially implicating California medical-licensing law.
• Whether claims against Altman survive as pleaded or are treated as corporate-conduct allegations improperly aimed at an officer.
• Whether the requested pause of ChatGPT Health receives emergency treatment, ordinary motion practice, or no early traction.
• How OpenAI uses the February 2026 retirement of ChatGPT-4o to distinguish the alleged 2025 conduct from current products.

Winters v. OpenAI is not precedent. It is not a causation finding. It is not yet a judicial answer to whether ChatGPT can be liable for dangerous medical advice. But as a filed complaint, it packages three theories that counsel now have to track together: AI software as a product, chatbot health guidance as possible unlicensed medical practice, and executive undertaking as a personal-liability route. That is enough to move it beyond the usual hallucination docket.

References

1. ChatGPT's medical advice nearly killed a Florida man, lawsuit against OpenAI claims, CBS News
2. Man sues OpenAI over dangerous medical advice from ChatGPT, Courthouse News Service
3. ChatGPT's advice kept man from seeking medical treatment for dangerous condition, lawsuit claims, Reuters
4. ChatGPT medical advice brought man to brink of death, lawsuit alleges, BBC
5. OpenAI Faces Product Liability Lawsuit Over ChatGPT Medical Guidance, Law Commentary