AI Injury Litigation Is Taking Shape

By David Meldofsky, Esq., Founder, Lawsuit Intelligencer

Winters v. OpenAI, No. CGC-26-639579, was filed August 3, 2026 in San Francisco County Superior Court, and the coverage has treated it as a ChatGPT medical advice case. The facts support that reading. A 55-year-old pastor asked a chatbot about his symptoms for roughly a year. It told him he was not yet in serious territory and should rest. He stopped moving, and in July 2025 he was hospitalized with clots in both lungs that his physicians attributed to immobility.

For firms deciding whether AI injury claims belong in their inventory, the facts are the least informative part of the filing. The signature block is where the useful information is. Read with the relief requested, it says this case is being built as precedent litigation, and firms should plan around that rather than around a settlement.

Who is on the caption

Three organizations represent the plaintiff. The Social Media Victims Law Center appears through Laura Marquez-Garrett, with Matthew Bergman pro hac vice. Tech Justice Law appears through Meetali Jain, Sarah Kay Wiley, and Tiffany Gillis Brown. The third is the Institute for Law, Innovation & Technology at Temple University’s Beasley School of Law, through Laura Bingham, also pro hac vice.

These three have run the same docket together before. The combination carried Garcia v. Character Technologies through its motion to dismiss in the Middle District of Florida. In May 2025, Judge Anne Conway declined to dismiss the core product liability and negligence claims at the pleadings stage, sending the product-versus-service question into discovery.

Garcia never produced an answer to that question. The parties told the court on January 7, 2026 that they had reached a settlement in principle, and four related family cases in Colorado, New York, and Texas were resolved at the same time. What it paid, and on what conditions, was never made public. Because it settled, no court has finally decided whether a chatbot is a product.

What the clinic tells you about the filing

Most readers will skim past the clinic line, or assume it means law students doing document review. Temple’s Institute for Law, Innovation & Technology runs a clinic where students handle real technology-harm cases, and it has worked alongside these same lawyers for years. In Garcia, the clinic helped write the opposition that defeated the motion to dismiss. A clinic on the caption means the plaintiff’s side has a standing research operation that does not bill by the hour, and the pleading shows it.

The complaint runs 62 pages and is built like a brief rather than a demand. It leans on published clinical literature, on OpenAI’s own safety disclosures, and on the departures of company safety researchers. Its most aggressive allegation is not about the plaintiff at all. The complaint cites a study that tested the model against hundreds of physician-written clinical scenarios and alleges that it missed high-risk emergencies in more than half of the acute cases, and that its warnings grew weaker as the emergencies grew more serious.

Those are allegations, untested, and OpenAI has not yet responded. But a complaint built this way is aimed past the defendant. It is written to survive a demurrer, then an appeal, and to leave behind an opinion the next hundred filings can cite.

What the complaint asks the court to do

This is the part with direct operational consequences, and it is the part that gets skipped.

The damages requests are ordinary. The injunctive requests are not. The complaint asks the court to make ChatGPT end a conversation automatically when someone needs immediate medical help, and to hard-code refusals on treatment and diagnosis questions so they cannot be worked around. It asks for warnings about dependency and dangerous medical advice, disclosure of what the model was trained on, and auditable data-provenance controls going forward. It asks the court to pause ChatGPT-4o Health, and any other consumer health product, until independent auditors clear it.

Then the two that matter most. The complaint asks the court to permanently destroy the GPT-4o model, or bar OpenAI from offering it to anyone for any purpose. And it asks for deletion of all training data and derivatives built from consumer use of that model.

There is nothing in that list for a settlement administrator to administer. The fee request points the same direction. Fees are sought under Code of Civil Procedure section 1021.5, the private attorney general statute, which pays counsel for winning a public benefit rather than for moving claims.

The theory that sidesteps the fight everyone is watching

The industry has spent two years watching three defenses: that a chatbot is a service rather than a product, that Section 230 immunizes the company, and that model output is protected speech. The fifth cause of action here does not depend on any of them. It pleads negligence under California Business and Professions Code sections 2052 and 4999.9, the unlicensed practice of medicine and unlicensed psychotherapy provisions. A court can reach that theory without first deciding whether a model is a product.

Two more counts are worth watching on demurrer regardless of what happens to the products claims. The seventh pleads negligent undertaking against Sam Altman personally, the theory that someone who volunteers to protect others takes on a duty to do it carefully. The eighth pleads invasion of privacy under Article I, Section 1 of the California Constitution.

Why a single-plaintiff case matters to inventory

Most AI injury cases to date have been death cases brought by families. Winters survived, and his injury is physical rather than psychiatric. That produces a claim profile firms already know how to work: ongoing specials, rehabilitation, wage and earning capacity loss, no probate, no wrongful death standing analysis, and a living plaintiff who can testify to reliance.

If this pleading survives demurrer, the survivor delayed-care claim stops being a novelty and becomes a recognizable claim type.

What this means for case workup

Six practical consequences for firms evaluating these claims.

Preserve the conversation history first. It is the primary evidence, and it sits in the claimant’s own account, on the defendant’s platform, where the claimant can delete it. Export instructions belong in the intake script, ahead of the retainer.

The export is not the whole file. Saved memories live apart from chat history, so deleting a conversation leaves the memory it produced untouched, and the standard data export does not reach the memory store at all. There is no bulk download for it. Someone has to open the settings surface and capture the entries by hand, and the mechanics have shifted more than once, so verify them at collection. That is not a technicality. A transcript records what the claimant typed. The memory store records what the product decided was true about him and applied every time he came back.

Look for the medical record that mentions the chatbot. In Winters, the treating physicians and therapists wrote down, at the time, that he was following ChatGPT’s advice instead of seeking care. That note was made before any lawyer was involved, and it is what connects the product to the injury. A claimant explaining after the fact that he trusted the chatbot is a different thing, and it will not do the same work. Ask at intake whether the chart says it.

Map dates of use to model versions. The relief sought here is specific to GPT-4o. Version, not brand, is likely to drive the liability theory and how cases get grouped, and intake forms that capture only “ChatGPT” will not support that later.

Do not plan around a settlement calendar. Garcia resolved in roughly fifteen months, but on terms no one outside the mediation saw, and the question it raised was never decided. A campaign staffed by a nonprofit and a university clinic is built to win a published ruling that binds the next case, and that kind of ruling comes out of an appellate court, not a settlement fund.

Watch coordination. Winters was filed in the same court that hosts In re ChatGPT Product Liability Cases, JCCP 5431, and the default there has been to stay newly noticed cases pending coordination. If Winters is drawn in, it may go quiet for months, and that will affect timing more than any other single procedural event.

What to watch

OpenAI’s response is the first real data point, and the question inside it is narrower than the headline: whether the Business and Professions Code claims survive on their own, apart from the products liability counts. If they do, plaintiffs have a route that does not run through the product-versus-service argument at all.

The second marker is whether the structural relief stays in the case as a live demand or turns out to be a bargaining position that trades away once damages are seriously discussed. The answer to that one tells you what kind of litigation this actually is.

This article is editorial analysis and does not constitute legal advice or an evaluation of any specific claim.