The First Chatbot Mass Tort Gets a Command Structure: Leadership Assigned in Nascent MDL

The First Chatbot Mass Tort Gets a Command Structure: Leadership, Liability Theories, and Pace in JCCP 5431

On February 3, 2026, the Superior Court of California, County of San Francisco, entered an order coordinating a group of California actions against OpenAI and its chief executive under the special title ChatGPT Product Liability Cases, Judicial Council Coordination Proceeding No. 5431.¹ Six months later, on August 4, 2026, the coordination judge, Hon. Ethan P. Schulman, entered the case management order that will actually determine how the litigation is run: an order appointing plaintiffs’ leadership.²

The contest that preceded that order deserves attention from anyone who defends product manufacturers or writes coverage for technology companies. In the days before the deadline, two camps of experienced plaintiff-side firms — one led by Edelson PC, the other by the Social Media Victims Law Center — were competing to run the case, under a court-imposed requirement that they submit a single joint proposal or let the court pick for them.³ They submitted a joint proposal. How a mass tort is organized at the outset shapes its discovery scope, its settlement posture, and its speed far more than most early merits rulings do.

A Coordination Proceeding, Not an MDL

The first point of orientation is jurisdictional. This is not a federal multidistrict litigation. It is a California state-court coordination proceeding under Code of Civil Procedure section 404 and rules 3.500 through 3.550 of the California Rules of Court, a mechanism that allows the Chair of the Judicial Council to assign civil actions sharing common questions of fact, filed in different superior courts, to a single coordination trial judge for all purposes.⁴ The practical effect resembles an MDL, but the differences matter: there is no mandatory remand-for-trial requirement of the sort Section 1407 imposes in federal MDL practice, so the coordination judge may keep the coordinated cases through trial, although rule 3.542 permits remand of a coordinated action or a severable issue to the originating court; California procedure — demurrers rather than Rule 12 motions, the discovery act rather than the Federal Rules — governs throughout.

The order swept in roughly a dozen actions originally filed in the Los Angeles, San Francisco, San Diego and Alameda superior courts, including Fox, Shamblin, Madden, Brooks and Gray from Los Angeles, Lacey, Enneking, Irwin, Raine and First County Bank from San Francisco, DeCruise from San Diego and Jacquez from Alameda. By the court’s own tally, five of the coordinated cases pleaded wrongful death and two involved minors.⁵ The named defendants across the coordinated actions include the OpenAI corporate entities and Sam Altman, along with Doe employees and Doe investors; reporting on the leadership order describes the coordinated litigation as running against OpenAI and Microsoft.⁶ The lead-off case, Raine, was filed in San Francisco in August 2025 and has driven much of the public framing of the litigation.⁷

Why the Leadership Fight Mattered

In a coordinated proceeding of this kind, the court appoints a plaintiffs’ leadership structure that controls the common-benefit work: master pleadings, defendant-side discovery, expert development, privilege fights, and negotiations over bellwether selection. Individual plaintiffs’ counsel retain their clients, but the leadership decides what theory the litigation will actually test. A structure assembled by agreement tends to be broader and more inclusive; one imposed by a court after a contested fight tends to be smaller and more hierarchical, and it leaves the losing camp with an incentive to relitigate leadership questions later.

Judge Schulman approved a jointly proposed structure consisting of four co-lead counsel, one liaison counsel, and a steering committee that MLex reports as having six members — and paired the appointment with an unusually explicit warning. The court, according to reporting on the order, cautioned plaintiffs’ counsel “that it expects counsel to work together collaboratively and without undue disagreement or delay, and to avoid unnecessary duplication of effort and expense,” and indicated it would restructure the leadership team if they failed to cooperate.⁸ That is a signal about the court’s tolerance for the internal friction that often slows large coordinated proceedings, and defense counsel should read it as such. A court that reserves the power to reorganize plaintiffs’ leadership mid-stream is a court that intends to keep the schedule moving.

The Liability Theories Actually on File

The claims are conventional in form and novel only in their object. In the publicly available pleadings — the Raine complaint and two other estate actions — plaintiffs plead strict liability for design defect and failure to warn, negligent design defect and failure to warn, violation of California’s Unfair Competition Law, wrongful death, and survival actions.⁹¹⁰ The theory is that a large language model deployed as a consumer product was designed in ways that made it foreseeably dangerous to vulnerable users — that safety classifiers degraded over long conversations, that engagement-oriented design encouraged emotional reliance, and that the company shipped the relevant model without adequate testing of those behaviors.

The doctrinal fulcrum is whether a generative model is a “product” at all for purposes of strict liability. California courts have historically been reluctant to treat pure information and software as products, and defendants can be expected to press that line hard. Plaintiffs’ answer is that they are attacking design choices — how the system was built, trained, and configured — rather than the informational content of any particular output. That distinction, borrowed from the social media addiction litigation, is the load-bearing move in the entire theory. If the court accepts it, strict liability and its consumer-expectations and risk-utility tests come into play. If it does not, plaintiffs are left with ordinary negligence, where duty, foreseeability, and the scope of any duty owed to a self-harming user become far more contestable.

Failure to warn is likely to be the most durable claim regardless of how the product question resolves. It does not depend on classifying the model as a product in the strict-liability sense, and it aligns with the factual record plaintiffs have built around what the company knew about model behavior in extended conversations. OpenAI’s own published safety discussion, which acknowledges that safeguards “can sometimes be less reliable in long interactions” as safety training degrades over an extended exchange, is the sort of contemporaneous corporate statement that plaintiffs will use on knowledge and notice.¹¹ That is a general lesson for technology clients: candid public safety disclosures are valuable and often necessary, but they are also discoverable admissions, and they should be drafted with litigation in mind.

Section 230, the First Amendment, and Causation

Three defenses will define the pleading stage. The first is Section 230 of the Communications Decency Act, which bars treating a provider of an interactive computer service as the publisher or speaker of information provided by another information content provider.¹² The immunity fits awkwardly here. A model that generates novel text is not obviously republishing content supplied by someone else, and plaintiffs frame their claims as attacks on design rather than on the transmission of third-party material. Expect the defense to argue that the outputs are shaped by user prompts and by training data authored by others, and expect plaintiffs to argue that generative output is first-party content that falls outside the statute entirely.

The second is the First Amendment. In the Character.AI litigation in the Middle District of Florida, the court declined at the pleading stage to hold that chatbot outputs are protected speech, reasoning that it was not prepared at that juncture to treat the outputs as expression.¹³ That ruling is not binding on a California superior court and it was expressly provisional, but it is the leading data point, and it means the constitutional defense is unlikely to dispose of the case at the demurrer stage. The more consequential question is whether the First Amendment ultimately constrains the remedies — whether an injunction that dictates model behavior differs constitutionally from a damages award for a design choice.

The third defense is causation, and it is the one most likely to matter at trial. OpenAI has publicly attributed the Raine death to misuse of the technology in violation of its terms of use, and has pointed to the decedent’s pre-existing risk factors, his circumvention of safety responses, and the many occasions on which the system directed him to crisis resources.¹⁴ Superseding cause, comparative fault, and the intervening act doctrine as applied to suicide are old problems in tort law with a substantial body of California authority behind them. What is new is the evidentiary texture: the conversation logs are a nearly complete record of the interaction, which cuts in both directions and will make these cases unusually document-driven at trial.

A Parallel Federal Track

Coordination did not consolidate everything. In April 2026, Chief Judge Richard Seeborg of the Northern District of California denied OpenAI’s motion to dismiss or stay a federal wrongful-death action under the Colorado River abstention doctrine, holding that although the federal case and a parallel San Francisco Superior Court action shared key facts and defendants, there was substantial doubt that the state proceeding would resolve the federal one.¹⁵ The court reasoned that the failure-to-warn analysis turned on the defendants’ knowledge of distinct risks — harm directed outward versus harm directed at the user — and that those inquiries were not necessarily coextensive.¹⁶

The result is two tracks proceeding at once, with overlapping defendants, overlapping causes of action, and no central coordinating authority between them. For the defense that means duplicative discovery, the risk of inconsistent rulings on the same product question, and a materially more complicated path to any global resolution. It also means that the first substantive merits ruling on AI product liability may well come from a federal judge in a single case rather than from the coordination court.

Bellwether Sequencing and the Question of Pace

With leadership in place, bellwether selection is the next structural fight. As of mid-August 2026, no bellwether trial date for the coordinated proceeding appears in the public record. Plaintiffs in mass torts routinely push for an early trial setting, and there is an obvious template: in the California social media addiction coordination, the first bellwether reached a jury in March 2026 and produced a verdict of roughly $6 million against two platform defendants; the trial court denied the defendants’ post-trial motions in June 2026, and both said they would appeal.¹⁷¹⁸ That result is a strong argument for speed from the plaintiffs’ side.

Take the summer of 2027 as a hypothetical benchmark, not a date set or proposed in any filing. A trial that soon would nonetheless be aggressive. The coordinated cases involve unresolved threshold questions of first impression, discovery into model training and safety evaluation that has no established protocol, and expert disciplines — machine learning safety evaluation, psychiatry, and causation in suicide — that have never been combined before a jury. Judicial caution about an early trial date in that posture would be unsurprising and, from a defense perspective, useful. The realistic assumption for planning purposes is that demurrer practice and the product-classification question resolve first, that bellwether workup follows, and that any trial date set now will move.

What Defense and Coverage Counsel Should Do Now

First, treat the leadership order as a scheduling signal rather than a personnel announcement. A court that appoints a broad structure and simultaneously threatens to restructure it is telling the parties it will not tolerate delay attributable to plaintiffs’ internal disagreements — and by implication will not extend patience to the defense either. Build discovery and expert timelines against a fast case, not a slow one.

Second, resolve the product-classification question early and in writing. Whether a model is a product for strict liability purposes is the single issue that most affects exposure, and it is worth briefing thoroughly at the demurrer stage even if the court is likely to defer it. Clients in adjacent industries — anyone deploying a conversational model in a consumer-facing setting, including insurers using chatbots for claims intake and health plans using them for member engagement — should understand that a ruling here will be cited against them within weeks.

Third, audit public safety disclosures and internal safety documentation with an eye toward discovery. The materials that will matter most are the ones companies produce voluntarily: safety blog posts, model cards, internal red-team results, and classifier threshold decisions. None of that should stop; safety work and its documentation are the best evidence of reasonable care. But it should be created with the understanding that it will be read aloud to a jury, and privilege boundaries around safety evaluation should be established before litigation, not after.

Fourth, map the coverage tower now rather than at tender. Claims of this type sit at an awkward intersection: bodily injury and wrongful death allegations point toward general liability, while the underlying conduct — software design, algorithmic output, professional-style advice — points toward technology errors and omissions. Many CGL forms contain professional services and electronic data exclusions that carriers will invoke; many tech E&O forms exclude bodily injury. Insurers have also begun adding express artificial intelligence exclusions and endorsements, and the scope of those provisions has not been tested. Claims naming a chief executive individually raise separate directors and officers questions, including whether allegations of pre-release pressure on safety testing are covered or fall within a deliberate-conduct carve-out.

Fifth, preserve and understand the conversation data. In these cases the interaction log is the central exhibit. Defense counsel should know, early, what is logged, how long it is retained, what the deletion and retention settings actually do, whether the data can be authenticated in the form it will be offered, and what privacy law constrains its use. Preservation obligations attach to model versions and configuration histories as well as to conversation text, and reconstructing which model behaved which way at which time is not something that can be done retroactively.

Sixth, watch the federal case as closely as the coordinated one. The parallel Northern District of California action is smaller, faster, and free of the leadership overhead, and it may generate the first appellate-quality ruling on whether these theories survive. A defense-favorable ruling there would reshape the coordinated proceeding; an adverse one would accelerate it.

The Larger Point

The leadership order does not decide anything about liability. What it decides is who will build the record, and in a mass tort that is nearly as consequential. The plaintiffs’ bar has now assembled, under judicial supervision, a coordinated apparatus aimed at establishing that a generative model is a product subject to design-defect law. Whether that theory succeeds will be worked out over the next two years in a San Francisco courtroom and, in parallel, in a federal one across the street. Every company deploying a conversational model in a consumer-facing role has a stake in the answer, and every carrier writing technology risk has a reason to read the pleadings now rather than after the first verdict.

This article was written by Arnold D. Lee, an attorney in the Phoenix, Arizona office of Spencer Fane. For more information, visit spencerfane.com.