Legal Guides

Is AI Output a Product or Content?

By David Meldofsky

Published June 12, 2026

Every major lawsuit against an AI company runs into the same question before anything else can happen: is what a chatbot says a product, the way a car or a medication is a product, or is it content, the way a book or a website post is content? The answer decides which body of law applies, which defenses are available, and in many cases whether the lawsuit survives its first motion. This guide explains the question in plain English, why it is genuinely hard, and how courts have started to answer it.

This is the doctrinal companion to our AI Lawsuits overview and the OpenAI case tracker. The full legal treatment of this question is David Meldofsky's Law360 Expert Analysis, Product-Or-Content Question Is Pivotal In AI Litigation, written for a professional audience. This page covers the same ground for non-lawyers.

Why One Question Controls Everything

The two classifications lead to two very different legal worlds.

If AI output is a product, its maker can be sued under product liability law. That is the framework courts use for cars, drugs, power tools, and medical devices. A plaintiff can argue the product was defectively designed, that safer alternative designs existed and were not used, or that the maker failed to warn users about known risks. Critically, none of those claims requires proving the company intended any harm. The focus is on the design choices the manufacturer made.

If AI output is content, two powerful defenses come into play. Section 230 of the Communications Decency Act protects online services from being treated as the publisher of information provided by someone else, and the First Amendment protects speech. Together, those defenses have ended most lawsuits against internet platforms for decades, usually before any evidence is gathered.

So when a family sues OpenAI over a death they connect to ChatGPT conversations, the first real fight is not about what happened. It is about which of those two worlds the case lives in.

Where the Line Came From

Product liability law grew up around physical goods, and information has always sat outside it. The classic case is Winter v. G.P. Putnam's Sons, a 1991 federal appeals decision about a mushroom encyclopedia. Readers who relied on the book ate poisonous mushrooms and were seriously injured. The court held that while the physical book might be a product, the ideas and information inside it were not, so strict product liability did not apply. Courts reached similar results for navigation charts and standalone software.

Software embedded inside a physical product has been treated differently. If the software in a car's braking system fails, the car is still a defective product. But those cases involved software bundled into a physical thing. Software sold on its own, like a chatbot subscription, has historically looked more like the encyclopedia than the car.

Why Generative AI Breaks the Old Categories

A chatbot's output is text, which sounds like content. But the system that produces the text is shaped end to end by its manufacturer: what data it was trained on, what behaviors were rewarded during training, what safety systems were layered on top, and what the company did when those systems flagged problems. Three features make AI harder to classify than any earlier software, and each one cuts both ways.

• The output is new every time. A chatbot generates its responses word by word rather than retrieving something pre-written. Defendants say that makes the output speech, not a manufactured item. Plaintiffs respond that the novelty is itself a design choice: the company built and tuned the system that generates it.
• The company did not write the specific words. No engineer at OpenAI typed the sentences a user sees. Defendants say that means the company is not the speaker. Plaintiffs respond that the company designed the system that produced the sentences, which is exactly how design-defect law works: nobody hand-builds each defective unit either.
• The behavior emerges from training. A company can know that harmful outputs will occur without being able to predict any specific one. Defendants say that makes specific harms unforeseeable. Plaintiffs respond that the category of harm was foreseeable, documented internally and externally for years, which is what foreseeability has always meant in tort law.

How Plaintiffs Frame It

The strongest version of the plaintiff argument is not that the chatbot said a bad thing. It is that the company designed, trained, and deployed a system whose foreseeable use produced specific categories of serious harm while safer alternatives went unused. That mirrors how plaintiffs plead defects in pharmaceuticals and medical devices: the defect is not the individual pill or sentence, it is the design decisions, the testing, the warnings, and the company's response once problems surfaced. All of that is manufacturer conduct, and manufacturer conduct is what product liability law has always evaluated.

The complaints in the current OpenAI docket plead several overlapping theories built on that framing, and some of them are designed to survive even if the product question goes against the plaintiffs:

• Defective design claims allege the model was trained to be agreeable in ways that made it affirm and follow vulnerable users, and that safer designs were available and not used.
• Failure of safety operations claims allege internal threat-detection systems flagged dangerous conversations and the company's response was inadequate. This theory asks only whether stated safety practices matched actual conduct, not whether the output is a product.
• Failure to warn claims allege the product was marketed as broadly safe, including for emotionally sensitive use, despite internal knowledge that it performed unreliably in exactly those situations.
• Negligent entrustment claims, with roots in cases about cars and firearms, allege the company kept providing access to a user it had reason to know was dangerous. The claim targets the access decision, not the content, which is why it may survive even a ruling that the output is protected content.

Section 230, in Plain English

Section 230 was written in 1996. Its core rule says an interactive computer service cannot be treated as the publisher of information provided by another person. It is the reason platforms generally cannot be sued over what their users post.

AI companies argue the rule fits chatbots too: the user's prompt is the information provided by another person, and the chatbot's answer is functionally a reply to it. Plaintiffs argue the opposite: with a chatbot there is no third party, because the system itself generates the words, and the statute was never meant to immunize a company from the consequences of its own system's design. Which reading wins is one of the central open questions in AI law, and the answer will shape the next generation of Section 230 cases well beyond chatbots.

The First Amendment Argument

Defendants also argue that algorithmically generated text is speech protected by the First Amendment, drawing on cases that recognized protection for data and computer code. Plaintiffs respond that holding a manufacturer responsible for a system's design is not the same as punishing a publisher for its message. Courts have so far been reluctant to resolve that question early in a case, which matters, because the longer it stays open, the more evidence plaintiffs can gather.

What the Garcia Ruling Did

The most consequential ruling so far came in Garcia v. Character Technologies, a Florida federal case brought by Megan Garcia after her 14-year-old son's suicide following extensive interactions with Character.AI chatbots. The defendants moved to dismiss on the three grounds described above: the chatbot is a service rather than a product, Section 230 applies, and the output is protected speech.

In May 2025, the court declined to dismiss the core product liability and negligence claims. The ruling did not hold that the chatbot is a product as a matter of law. It held the plaintiff had plausibly alleged it was, that the conduct described could be characterized as design rather than publishing, and that the First Amendment question could not be decided at the pleadings stage. That is a procedural ruling, not a final one. But it was the ruling plaintiffs needed, because it sends the question to a developed factual record, where internal documents, training decisions, and safety records come into evidence. The full case guide is at Garcia v. Character Technologies, with the broader docket in our Character.AI lawsuits overview.

How the OpenAI Cases Test the Question

The current OpenAI docket approaches the question from several angles at once, which distributes the risk that any single theory fails. Raine v. OpenAI carries the design-defect theory on its strongest alleged facts. The Tumbler Ridge school shooting suits carry the safety-operations theory, which does not depend on the product question at all. The FSU shooting case carries negligent entrustment in its most traditional form. The Turner-Scott overdose case carries failure to warn in a medical information setting. And Florida's June 2026 civil enforcement suit shows state regulators pressing the same threshold question through consumer protection law; that suit is covered at Florida v. OpenAI, with the wider enforcement docket tracked at States Suing AI Companies.

Motion-to-dismiss rulings expected through late 2026 will begin to answer the question. No single ruling will settle it. The pattern across the rulings, which theories survive and which courts adopt which framing, is what will define the terrain.

Common Questions

Why does it matter whether AI output is a product or content?

Because the classification decides which body of law applies. Product status opens the door to design defect and failure-to-warn claims, the framework used for cars and drugs. Content status opens the door to Section 230 and First Amendment defenses, which have historically ended platform lawsuits before evidence is heard.

Has any court decided whether an AI chatbot is a product?

Not as a matter of law. Garcia v. Character Technologies held in May 2025 that a plaintiff had plausibly alleged a chatbot was a product, which allowed the case to proceed to discovery. That is the furthest any court has gone, and it is procedural rather than final.

Does Section 230 protect AI companies?

It is unresolved. The defense reading treats the chatbot's answer as a reply to user-provided information. The plaintiff reading says the system itself generates the words, so there is no third party to point to. Courts are testing both readings now.

Can a case succeed even if courts call the output content?

Possibly. Negligent entrustment, failure to warn, and safety-operations theories each target company conduct rather than the content itself, and each may survive an adverse ruling on the product question.

Sources and further reading

• David Meldofsky, Product-Or-Content Question Is Pivotal In AI Litigation, Law360 Expert Analysis (June 2026)
• Lawsuit Informer: OpenAI Lawsuits Hub and Case Tracker
• Lawsuit Informer: Raine v. OpenAI Case Guide
• Lawsuit Informer: Character.AI Lawsuits
• Lawsuit Informer: States Suing AI Companies
• Lawsuit Informer: Product Liability Lawsuits

Educational commentary only. Not legal advice. No attorney-client relationship is created.