The State of California has already answered the question in a caption. The coordinated proceeding gathering the lawsuits against OpenAI in San Francisco Superior Court is formally titled In re: ChatGPT Product Liability Cases, JCCP No. 5431. The word doing all the work in that title, “product,” is the single most contested word in AI law right now, because whether a large language model is a product decides who has to prove what, and therefore who pays.
If ChatGPT is a product, plaintiffs can plead strict liability: show a defect and causation, and the manufacturer’s care, diligence, and good intentions become irrelevant. If it is a service, or speech, plaintiffs must prove negligence, a duty owed and breached, which is a far harder case against a company that publishes safety frameworks and ships guardrails. Two years ago the question was academic. Today it is being litigated simultaneously in San Francisco and Tallahassee, and, by the signal of an August 11 filing in Leon County, it is heading into Florida’s state courts too.
The definition sounds like semantics. It is the entire architecture of AI liability, and the first courts to answer it are trial courts answering it right now.
Why one word carries the whole case
American products liability law grew up around tangible things. The Restatement (Third) of Torts defines a product as tangible personal property distributed commercially, and courts spent three decades holding that information is not that. The canonical case is Winter v. G.P. Putnam’s Sons, 938 F.2d 1033, where the Ninth Circuit in 1991 refused to treat The Encyclopedia of Mushrooms as a product after readers were poisoned following its advice. Ideas and expression, the court held, are not subject to strict liability, or authors would stop writing. The Third Circuit extended the logic to algorithms in Rodgers v. Christie in 2020, holding that a pretrial risk assessment score was neither tangible nor the kind of thing products law was built for.
That wall of authority is what every AI defendant stands behind, and it held for decades against video games, search engines, and social feeds. Then, in May 2025, a federal judge in Orlando put the first crack in it.
Garcia: the ruling that moved the wall
Megan Garcia sued Character Technologies, its founders, and Google in October 2024 after her fourteen-year-old son died by suicide following months of immersive conversations with Character.AI chatbots. The defense ran the three arguments that now appear in every AI liability case: that a chatbot is a service rather than a product, that Section 230 immunizes the platform, and that model outputs are protected speech under the First Amendment.
On May 21, 2025, Judge Anne C. Conway denied the motion to dismiss in substantial part, in an opinion now reported at 785 F. Supp. 3d 1157. The strict product liability, negligence, and wrongful death counts went forward. On the First Amendment, the court declined to treat model output as protected speech at the pleading stage, reasoning that words assembled by probabilistic determination lack the human intention that expression requires, an analysis that drew on Justice Barrett’s concurrence in Moody v. NetChoice. The service-versus-product argument, the industry’s forty-year shield, failed to end the case at the threshold for the first time.
Then something more revealing happened than any ruling: nobody appealed the question. In a court filing on January 7, 2026, the parties disclosed a mediated settlement in principle covering Character Technologies, its founders, and Google, resolving four related family cases in Colorado, New York, and Texas at the same time. Garcia produced no merits judgment and binds no other court. What it produced instead is a playbook and a price signal. The playbook: plead the model’s design choices as defects and force the definitional fight past the pleadings into discovery, where internal documents live. The price signal: faced with litigating whether their system is a defective product on a full factual record, the defendants paid. Every plaintiff’s firm in the country read both.
The front that opened next
The suit that carried the theory to OpenAI is Raine v. OpenAI, filed August 26, 2025 in San Francisco Superior Court, case CGC-25-628528, by the parents of a sixteen-year-old who died by suicide in April 2025 after months of ChatGPT conversations. The complaint pleads strict liability and negligence against OpenAI and its chief executive, alleging design choices, sycophantic engagement, anthropomorphic memory features, and safeguards that weaken over long conversations, made the product unreasonably dangerous. The Raines amended in October 2025 to add intentional misconduct and punitive damages. OpenAI answered in late November 2025, denying causation and arguing the harm resulted from misuse of the product in violation of its terms of use, an answer the family’s counsel publicly called disturbing. The case sits in discovery and motion practice inside JCCP 5431, no trial date set.
The front widened on a schedule. In November 2025, seven families sued OpenAI over a school shooting in Canada. Estates in Connecticut sued over a murder-suicide in which the perpetrator had spent hundreds of hours with ChatGPT. A May 2026 suit alleges the model advised a fatal drug combination. And then Florida became the most aggressive jurisdiction in the country.
On April 17, 2025, a gunman opened fire at Florida State University’s student union, killing two people and injuring several others. The accused shooter, who has pleaded not guilty and faces trial in October 2026, allegedly used ChatGPT extensively in the months before the attack. In April 2026, after reviewing the chat logs, Florida Attorney General James Uthmeier opened a criminal investigation into OpenAI, the first anywhere, saying that if ChatGPT were a person, it would be facing charges for murder. On May 10, 2026, the family of Tiru Chabba, one of the two people killed, filed a federal wrongful death suit in the Northern District of Florida against OpenAI and the accused shooter. The complaint alleges thousands of messages in which the model discussed firearms operation and campus timing, and claims the system either failed to recognize an escalating threat or was not designed to. OpenAI’s public response is that ChatGPT provided factual, publicly available information and did not encourage or promote harm.
Which brings the story to a quiet docket entry. On August 11, 2026, a petition was filed in Leon County Circuit Court, captioned Mall v. OpenAI, seeking approval of a modified attorneys’ fee contract for product liability claims arising from the FSU shooting, filed by Morgan & Morgan. A fee petition is not a complaint. It discloses no theory, names no defect, and proves nothing about the merits. What it signals is procedural and specific: a second legal team, for a second set of claims connected to the same shooting, is formalizing its engagement for a product liability case in Florida state court, where the federal software precedents bind no one. Counsel for the family of the second person killed had already said publicly they intended to sue. The Florida front is becoming a two-court problem.
Design is an engineering word, and that is the plaintiffs’ whole theory
The reason the product frame has traction against LLMs, where it failed against books and video games, is that an LLM’s behavior is the output of enumerable, versioned, testable engineering decisions. A design defect claim needs a design. ChatGPT has one, and OpenAI documents it.
The training corpus is curated. The reward models that shape the system during reinforcement learning from human feedback encode explicit preferences about how agreeable, persistent, and emotionally engaged the model should be. The system prompt is a written instruction set. Refusal behavior is governed by classifiers with tunable thresholds. The Model Spec is a published design document stating intended behavior. Memory and session length are product decisions with measurable behavioral consequences. None of this resembles an author writing a mushroom encyclopedia. It resembles a manufacturer setting tolerances, which is precisely the analogy plaintiffs need.
Two episodes from OpenAI’s own publications hand the theory its best exhibits. In late April 2025, OpenAI shipped a GPT-4o update, then rolled it back within days after the model became conspicuously sycophantic, and published a postmortem attributing the failure to training adjustments that overweighted short-term user feedback. That is a manufacturer describing a behavioral defect introduced by a design change and corrected by a subsequent one, in public, with a root cause analysis. Then on August 26, 2025, the day the Raine suit was filed, OpenAI published a safety post acknowledging that its safeguards are more reliable in short exchanges and that in long conversations parts of the model’s safety training “may degrade.” The central technical allegation in Raine, that protections erode exactly when a vulnerable user needs them most, appears in the defendant’s own words.
The subsequent fixes keep writing the plaintiffs’ feasibility brief. In September 2025, OpenAI announced routing of sensitive conversations to reasoning models and parental controls. In December 2025 it shipped teen-specific Model Spec rules, an age-prediction system, and real-time classifiers for distress. Each change is a concrete alternative design that existed within months of the alleged harms, which is how design defect cases measure feasibility. Evidence rules complicate this, courts generally exclude subsequent remedial measures as proof of defect, with exceptions when feasibility is disputed, but discovery does not need the fixes themselves. It needs the internal experiments, threshold configs, and eval results that preceded them.
The strongest case against the product frame
Honesty requires the counter-mechanism, and it is substantial. A product, in the strict liability sense, is a fixed artifact replicated identically: every defective lawnmower is the same lawnmower. An LLM’s output is stochastic and conditioned on the user’s own inputs. No two conversations are the same unit, and the harmful transcript exists only because a specific user steered a probability distribution across thousands of turns. That is a genuinely hard fit for a doctrine built on identical widgets, and it is the intellectually serious version of OpenAI’s misuse defense.
Causation is harder still, and hardest in Florida. In the self-harm cases, the defense argues intervening personal crisis. In the FSU case, the causal chain runs through a deliberate criminal act by a third party, which the law ordinarily treats as superseding. The complaint tries to pre-empt this by pleading that the model’s own conduct, allegedly discussing logistics and reinforcing intent across thousands of messages, made the crime foreseeable to the system’s operator. Whether any court will stretch foreseeability across an intervening mass shooting is the steepest question on this entire front, and OpenAI’s publicly-available-information defense is the Winter argument reborn: what the model supplied was information, and information is not a product.
Section 230 and the First Amendment remain live too. Garcia rejected both at the pleading stage, but a pleading-stage denial in one district binds nobody, and no appellate court has touched either question for generative outputs. The industry’s wall lost one skirmish. It has not been demolished.
Technical claims audit
Grading the front’s load-bearing technical claims against documented system behavior, the standard this series applies to every case.
Safeguards degrade over long conversations. Consistent. OpenAI’s August 2025 post says so in nearly those words, and the mechanism is technically coherent: safety behavior instilled in post-training competes with in-context signals, and as context grows, the balance shifts. This is the single most consequential admission on the docket because it converts a plaintiff’s inference into a defendant’s statement.
Sycophancy is trained behavior, not anomaly. Consistent. The April 2025 rollback and postmortem document a causal chain from a reward-signal change to model behavior and back. Whatever a jury eventually makes of it, the incident establishes that engagement-adjacent behavior is a tunable design parameter, which is the factual predicate the design defect theory requires.
The system could have detected and escalated threat patterns. Contested. The FSU complaint’s failure-to-escalate theory assumes classification of dangerous intent across thousands of messages was feasible at 2024 to 2025 scale. OpenAI shipped real-time distress classifiers in December 2025, which proves feasibility now, but false-positive rates at hundreds of millions of weekly users are a genuine engineering constraint, and what was achievable when matters legally. Internal eval records will settle this. Public information cannot.
Safety testing for GPT-4o was compressed. Contested. The Raine pleadings and surrounding reporting allege a rushed evaluation timeline under competitive pressure. Schedules, sign-offs, and red-team reports are discoverable documents, and this claim will resolve into dates on a calendar.
ChatGPT provided only publicly available information. Underspecified, and this is OpenAI’s claim, not the plaintiffs’. Availability is not the same property as synthesis, personalization, or sustained engagement. A library contains what a research assistant does, and no one confuses the two. Whether the released logs show retrieval or show something more interactive is exactly what the Florida criminal investigation and civil discovery will establish, and the defense’s strongest sentence will live or die on transcript excerpts.
What discovery would settle it
The artifacts that decide this front are internal and specific: pre-launch safety evaluation suites and their results for GPT-4o, red-team reports and the decisions made on them, classifier threshold configurations and their change logs, any experiments measuring engagement against safety interventions, the Model Spec’s version history, and internal incident trackers for self-harm and violence-adjacent conversations. In the FSU matter, add the complete chat logs, portions of which law enforcement has already released, and the model-version and safety-configuration records for the exact period of the alleged conversations. Every one of these is a document that exists. Product litigation is the legal system’s only mechanism for compelling their production, which is, candidly, half the reason this front exists.
The ruling to watch
Two courts hold the trigger. In California, the first dispositive ruling inside JCCP 5431 on whether ChatGPT is a product for strict liability purposes will be the most consequential sentence in AI law to date, and unlike Garcia, it will issue from the coordinated proceeding every subsequent California case joins. In Florida, watch the motion to dismiss in the federal Chabba case and, if the Leon County signal matures into a filed complaint, the first state court demurrer. Florida recognizes strict products liability but has never held software to be a product, and its courts owe the Ninth Circuit’s Winter line nothing. A California yes and a Florida no, or the reverse, creates exactly the kind of split that forces legislatures and, eventually, higher courts to answer what the trial courts started.
One more possibility deserves naming because Garcia proved it: settlement as doctrine management. Every defendant on this front has now watched the product question survive a motion to dismiss once and get paid off before it could be answered on the merits. If OpenAI settles the cases that pose the question most sharply and litigates only the ones with the weakest causation, the definitional question stays open indefinitely while the practical answer, that these claims have settlement value, gets written into every insurer’s model. Watching which cases resolve quietly is as informative as watching which ones produce opinions.
Limits
Everything alleged remains allegation. No court has found any AI system defective, no appellate court has held a model to be a product, and OpenAI denies wrongdoing in every matter described here. The accused FSU shooter is presumed innocent, and his criminal trial has not occurred. The Leon County petition discussed above is a fee filing whose underlying complaint, if one is ever filed, has not been read by this publication, and no characterization of its claims beyond the docket’s own description is offered. Garcia settled without any merits determination, and its pleading-stage reasoning binds no other court. This piece analyzes a legal front in motion, and the record will move under it.
Frequently asked questions
Is ChatGPT legally a product?
Unresolved. Decades of precedent, including Winter v. G.P. Putnam’s Sons and Rodgers v. Christie, treat information and algorithms as outside product liability law. In May 2025, a federal judge in Garcia v. Character Technologies allowed strict product liability claims against an AI chatbot maker to proceed past dismissal, the first ruling of its kind, but the case settled before any merits decision. California’s coordinated ChatGPT proceeding, JCCP 5431, is where the question is most likely to be answered next.
What did the Character.AI ruling actually decide?
Judge Anne Conway’s May 21, 2025 order denied most of the motion to dismiss: product liability, negligence, and wrongful death claims proceeded, and the court declined to hold at the pleading stage that model outputs are First Amendment speech. It was a procedural ruling about the sufficiency of allegations, not a finding that Character.AI was defective, and the case settled in January 2026.
Does Section 230 protect AI chatbot outputs?
Untested at any appellate level. Section 230 shields platforms from liability for third-party content, and courts including the Garcia court have been skeptical that model-generated text, which no third party authored, fits that shield. The question has not been definitively resolved anywhere.
What is Raine v. OpenAI?
The foundational wrongful death suit against OpenAI, filed August 26, 2025 in San Francisco Superior Court by the parents of a teenager who died by suicide, pleading strict liability and negligence over ChatGPT’s design. OpenAI answered in November 2025 denying causation and citing misuse of the product. It is in pretrial litigation inside the coordinated proceeding as of this writing, with no trial date set.
Court records cited by caption, court, county, and filing date were retrieved from a commercial index of state trial court dockets on August 16, 2026. Federal case details are drawn from public filings and contemporaneous reporting linked above. Allegations are exactly that, and nothing here is legal advice. If this coverage touches on difficult experiences, support is available: in the US, call or text 988. This piece is part of an ongoing series on AI cases in state courts. The series opener on why AI lawsuit counts are broken and the companion analysis of whether an AI phone agent is a wiretap cover the wider docket.