The Chatbot Is Not One Legal Thing.

The allegations are grave. They are still allegations

Reuters reported on October 7 that plaintiffs have brought a cluster of cases alleging that interactions with ChatGPT contributed to suicides, an attempted overdose and mental-health crisis, a school shooting, and a murder-suicide. The cases differ in their facts, parties, jurisdictions, and theories. Several assert defective-design or failure-to-warn claims. OpenAI denies that it is liable and has begun making substantive arguments for dismissal.

The procedural posture matters. In Tallahassee litigation arising from the Florida State University shooting, OpenAI filed a motion to dismiss on September 25. California cases have entered coordinated discovery, Reuters reports, but no trial dates have been set. OpenAI's responses to several San Francisco complaints are due later this month. None of those facts amounts to a judicial finding that ChatGPT caused an injury, that a model was defective, or that OpenAI breached a legal duty.

The complaints deserve the same precision. The suit filed by the executor of Suzanne Adams's estate alleges that GPT-4o reinforced Stein-Erik Soelberg's paranoid delusions before he killed Adams and himself. The complaint describes product decisions, model behavior, and internal safety choices, but those assertions are a plaintiff's allegations. They have not been tested through discovery or proved in court. Tragedy does not relax the burden of proof, and a defendant's categorical description of its technology should not decide the case before that proof can be developed.

Reuters reports that most of the pending claims involve GPT-4o, a model introduced in May 2024 and retired in February 2026. That detail is not incidental. AI systems change through model updates, system instructions, classifiers, interfaces, and safety interventions. A legal claim about one configuration at one time cannot automatically establish anything about every version called ChatGPT.

The noun changes the lawsuit

Product liability offers plaintiffs an important path because it can focus on the condition of a product rather than only on the reasonableness of a conversation. Depending on the jurisdiction and theory, plaintiffs may argue that a design was defective, that foreseeable risks required a warning, or that a safer feasible design was available. Negligence claims ask related but distinct questions about duty, breach, foreseeability, and reasonable care. Neither framework makes liability automatic.

OpenAI's Florida motion argues that ChatGPT is an interactive service, not a product, and that information and ideas are not products merely because software delivers them. It also invokes constitutional protection for disseminating factual information. Those arguments target the threshold. If the court accepts the broadest version of them, some product claims could end before discovery reaches model design, safety testing, user signals, or internal decisions about deployment.

Plaintiffs answer that the alleged defect is not simply a bad idea expressed in text. Their complaints target a designed system that generates individualized responses, remembers context, responds continuously, and allegedly behaves in predictable ways under certain conditions. The distinction matters. A book gives the same words to every reader. A conversational model produces new output through software operating on a user's prompts, prior exchanges, model parameters, and product-level instructions.

That does not prove the software is a product under any particular state's law. Product definitions vary, and courts have long struggled with software, data, and intangible information. It does show why analogy by noun is inadequate. Calling the system a service emphasizes the continuing relationship. Calling it a product emphasizes design and distribution. Calling its output speech emphasizes expression. All three observations can be true without answering which legal rule governs the particular conduct alleged.

The whole chatbot is the wrong unit of analysis

Courts should begin with the claimed defect, not the brand name. Does the plaintiff challenge the substance of an answer, the absence of a warning, a model's tendency to agree with a user, the system's memory of prior exchanges, the design of engagement loops, an age gate, a crisis-detection classifier, an escalation protocol, or the decision to deploy a particular model? Those are not interchangeable acts. They may implicate different duties and defenses inside the same application.

California's published decision in Hardin v. PDX offers a useful, limited analogy. The court distinguished allegedly defective software that reprogrammed prescription instructions from the information distributed through a database. At the pleading stage, it would not reject the software theory merely because information itself is not a product. Hardin did not decide the status of generative AI, and its holding does not govern every jurisdiction. Its method is still sound: separate the operation of software from the information the software conveys before deciding that an information rule resolves both.

A feature-level approach also prevents the opposite overreach. Plaintiffs should not be able to convert every offensive, false, or harmful sentence into strict product liability by describing the entire application as a defective product. A claim directed at an idea or message raises different concerns from a claim that a safety control failed to activate despite specified signals. The relevant unit may be a feature, a model version, or a deployment decision, not every output and not the entire service.

This is familiar work for courts. A newspaper can be a business, a physical product, and a vehicle for protected speech. A medical device can contain software, instructions, data, and professional services. The law does not need one metaphysical answer about what the enterprise really is. It needs an administrable account of the conduct for which liability is sought.

Speech and design are different defenses

The First Amendment issue is serious. Liability aimed at the substance of generated language can burden expression, and courts should be wary of theories that impose damages because a system communicated an unpopular idea or accurate public information. OpenAI's motion argues that the plaintiffs' claims impermissibly target protected speech. Plaintiffs argue that algorithmic output is not speech or that their claims concern product design rather than expression.

The early Character.AI litigation shows how little has been finally decided. In Garcia v. Character Technologies, a Florida federal judge declined at the pleading stage to hold that large-language-model output was protected speech and allowed most claims to proceed. The order did not find the chatbot defective, decide causation, or establish a general rule that AI output lacks constitutional protection. The case later settled. It demonstrates that a categorical speech defense may not resolve fact-intensive design allegations before a record exists.

The better boundary tracks the object of regulation. A claim that liability follows from what an answer means sits close to expression. A claim that a company omitted a neutral warning, disabled a crisis-response feature, or selected engagement settings despite documented risks may concern conduct and architecture even though words appear on the screen. Real cases will mix the two. Courts should identify the burden on expression rather than treating the presence of language as an all-purpose shield or the presence of code as a reason to ignore speech.

Classification only opens the courthouse door

Even if a court permits a product theory, plaintiffs must still prove a legally cognizable defect, the applicable standard, and causation. These cases involve human decisions, mental-health conditions, third-party acts, long sequences of interaction, and changing model behavior. Defendants will contest foreseeability, intervening causes, user misuse, the adequacy of warnings, and whether the alleged alternative design would have prevented the injury. State law may frame those questions differently.

Discovery will therefore matter more than rhetoric about whether AI is unprecedented. The useful record will include the model version involved, system and safety instructions, relevant evaluations, known failure modes, incident reporting, post-deployment monitoring, user-facing warnings, escalation rules, changes made before and after the events, and the evidence connecting a challenged feature to the alleged harm. Broad demands for every internal AI document will be expensive and imprecise. A feature-level liability theory should produce feature-level discovery.

Defendants need comparable precision. A platform should not be able to call the system a service when resisting product liability, speech when seeking constitutional protection, and a neutral tool when disputing responsibility without explaining which function each characterization describes. Those arguments may all be legally available. Their force depends on disciplined boundaries rather than strategic labels.

The hardest part of these cases will not be choosing between product, service, and speech. It will be deciding which design choices are properly attributed to the developer, which risks were reasonably foreseeable at the time, whether a duty reached the particular plaintiff, and whether admissible evidence can connect the choice to the injury. Classification determines whether that inquiry occurs. It should not be mistaken for the inquiry itself.

Do not grant immunity by taxonomy or liability by tragedy

Courts now face two seductive shortcuts. One is immunity by taxonomy: the system produces words, words are information or speech, and therefore claims about its operation cannot proceed. The other is liability by tragedy: the alleged harm is devastating, the interaction was intimate, and therefore the developer must answer in damages. Neither approach does the legal work.

A conversational AI system is not one legal thing because it does more than one thing. It delivers a service, operates software, makes design choices, and generates language. The proper classification can change from claim to claim without inconsistency. What matters is whether the court identifies the alleged conduct, applies the governing state's law, protects expression where expression is actually burdened, and requires proof of defect and causation where a product theory survives.

These lawsuits may ultimately fail, settle, or produce sharply different results across jurisdictions. At this stage, the soundest position is modest but consequential: no party should win merely by selecting the noun for the technology. The law should classify the function before it assigns the consequence.

A court does not need one universal answer to what a chatbot is. It needs a precise answer about what the plaintiff says went wrong.