Who Answers for What a Machine Says and Does?

Not so

On June 1, 2026, Florida’s Attorney General sued five OpenAI companies and their chief executive, Sam Altman, in the circuit court for Highlands County. The first page of the State’s complaint (opens in a new tab) reprints the promise ChatGPT shows its users under the heading “Built with safety in mind.” Safety, it says, is built into how the company designs, tests, and improves the product. Below it, the State answers in two words: “Not so.”

Those two words open an argument that tort law has to finish. The law of accidents grew up around harms with an author. A chatbot’s answer has several: the company that trained the model, the company that offered it to the public, the executive who, the State says, directed both, and the user who typed the question. So who answers? This reading follows the Florida complaint through three of its ten counts, negligence and the two product claims. Then it takes up two questions the complaint never raises: what happens when a chatbot says something false about a real person, and when software acts instead of talking. Those are the questions behind the two chains and the statute in the activity.

Five companies and a chief executive

A liability chain lists everyone a plaintiff could reach, with the theory that reaches each one and the defense each would raise. The Florida complaint names six defendants and does not sort them. Five are OpenAI companies with different jobs on paper. One is a nonprofit parent, one a public benefit corporation that builds and distributes the product, and the others an operating company, its parent, and a holding company that owns the intellectual property. Then the complaint says that together they “own, operate, control, produce, design,” and distribute ChatGPT (¶ 16). In this record the developer and the deployer are one family of companies, and the State treats them as one.

The sixth defendant is a person. The complaint reaches Altman through what it says he did, not through his title: he “personally directed the design, development, safety policies, and deployment” of ChatGPT (¶ 15).

Other chains are longer. In Garcia v. Character Technologies, Inc. (opens in a new tab), a wrongful-death suit in the federal court in Orlando, the plaintiff sued the company that ran the Character.AI app, its two founders, and Google. Google had not run the app. The complaint alleged that it contributed technology to the app’s design, and Judge Anne Conway let the claim go forward on a component-part theory (pages 18 to 20). Under that theory, the maker of a part can answer for the finished product when it substantially took part in building the part into the design, and doing so made the product defective. A developer can be reached through a deployer’s product.

Count V, the negligence count, alleges that each defendant had a duty to use appropriate care in designing, marketing, and selling ChatGPT (¶ 243). Florida decides whether a duty exists by asking whether the defendant’s conduct created a foreseeable zone of risk to others. In McCain v. Florida Power Corp. (opens in a new tab), the Florida Supreme Court called duty “a minimal threshold legal requirement for opening the courthouse doors.” The harder fight comes later, at proximate cause, where the plaintiff must show that the conduct foreseeably and substantially caused the specific injury. Garcia applied the rule to a chatbot. By releasing the app to the public, the court held, the defendants created a foreseeable risk of harm that they were in a position to control (page 38).

That last phrase matters for the third link. Nobody in these records sues the user. But the user’s choices are where a defendant looks for its answers: the warnings the user saw, the uses the user chose, the control the user had. The Florida complaint sees the argument coming. It alleges a duty to warn Floridians of dangers even when ChatGPT is “not used for its intended purpose” (¶ 282).

Is a chatbot a product?

Counts VII and VIII put ChatGPT into the law of products, and the choice matters. Strict liability looks at the product rather than at the maker’s care. If the product was defective when it left the maker’s control and the defect caused the harm, a careful maker still answers. The complaint puts it in one line: “ChatGPT is a product subject to Florida’s strict product liability laws” (¶ 262).

Florida’s test for a design defect comes from Aubin v. Union Carbide Corp. (opens in a new tab). A design is defective if the product fails to perform as safely as an ordinary consumer would expect when used as intended or in a reasonably foreseeable way. The Florida Supreme Court refused to adopt the Restatement (Third) of Torts, which would make the plaintiff prove that a reasonable alternative design existed. One reason it gave: a maker shapes what consumers expect of its product. The complaint pleads the consumer-expectations test almost word for word (¶ 264). It pleads alternative designs too (¶¶ 270 to 272), which Florida allows but does not require.

Before any of that, a court has to decide whether a chatbot is a product at all. Courts have generally refused to treat ideas, words, or expression as products, so the story in a video game does not bring strict liability. Page 33 of the Garcia order collects the cases. Judge Conway drew the line inside the app. Some of the claims in Garcia pointed at features of the app, such as the lack of any age check. The court held the app a product “so far as” the claims arise from defects in the app “rather than ideas or expressions within the app” (page 36). A claim about how a chatbot was built can cross that line. A claim about what it said may not. Count VII’s allegations fall on both sides, and sorting them is part of your chain.

Failure to warn, the second product claim, asks a narrower question. Count VIII alleges that users were never warned of the risks the complaint describes (¶ 285), and the complaint’s first page sets a promise of safety where a warning might have been. A defendant answers that its terms and on-screen notices did warn. Whether a warning was adequate, and whether it reached the person harmed, depends on facts, and courts tend to wait for them. The Delaware court in the next section refused to weigh Google’s disclaimers without a record of who saw which one.

When the machine talks about someone

The Florida complaint never mentions Section 230 or the First Amendment. A company sued for what its chatbot said about a real person meets both, and the first round of one such case is on the record.

Robert Starbuck, a filmmaker and activist who lives in Tennessee, sued Google in Delaware over twenty-two statements its AI tools made about him in 2025, which he says were fabricated. Google did not argue that they were true. It moved to dismiss for want of damages, publication, and actual malice, and in July 2026 the court in Starbuck v. Google LLC (opens in a new tab) denied the motion in its entirety.

Publication and fault carried the case. On publication, the court asked whether anyone besides Starbuck received the statements. He alleged that his children and colleagues did, that two people asked him about them, and that Google’s AI admitted giving false statements about him to nearly three million users. That was enough for now. The court did not decide who “publishes” words a model generates. Google had assumed for the motion that AI output can be defamatory, and the court noted “the murky waters of defamation law in this context” (page 13, note 60). On fault, the parties agreed for the motion that Starbuck is a public figure, so he must prove actual malice: that Google knew the statements were false or recklessly disregarded whether they were. Part of his answer was notice. He alleged that his letters reached Google’s legal department before the statements he sued on.

Two defenses waited for a record. Google pointed to disclaimers telling users its AI can be wrong, and the court would not decide on the pleadings what disclaimer anyone saw: “A more robust record is necessary” (page 21, note 93). Google also relied on Walters v. OpenAI, where a Georgia court entered summary judgment for OpenAI in 2025. The Delaware court found that ruling inapposite, because it came after discovery (pages 19 and 20). Florida would ask the same questions through the five elements in Jews for Jesus, Inc. v. Rapp (opens in a new tab): publication, falsity, fault, damages, and a defamatory statement.

Section 230 appears nowhere in the Starbuck opinion. The statute protects an online service from being treated as the publisher of “information provided by another information content provider.” In Zeran v. America Online, Inc. (opens in a new tab), the Fourth Circuit read it to bar a suit over a user’s defamatory posts. But the statute defines an information content provider as anyone “responsible, in whole or in part, for the creation or development” of the information. When a company’s own model writes the sentence, the company is at least partly responsible for it. In other words, the words may not be “provided by another” at all. That reading of the text is why the defense looks harder for a chatbot than for a message board, though courts have not settled it.

The First Amendment defense asks something else: whether a model’s output is protected speech. In Garcia, the defendants said it was. Judge Conway wrote that they “fail to articulate why words strung together by an LLM are speech” and declined to hold the output speech “at this stage” (pages 28 and 31). She did not hold that it is unprotected either. The question is open.

When the machine acts

Everything so far concerns what a machine says. Software that acts, buying, booking, or filing for its user, raises an older question, and contract law has an answer. Florida’s Uniform Electronic Transaction Act defines an electronic agent as a program used on its own to start an action or respond to records without review by a person (section 668.50(2)(f)). A contract can form through one “even if no individual was aware of or reviewed” what it did (section 668.50(14)(a)). In other words, a person who sends an agent to make a deal cannot escape the deal by pointing out that no human was watching.

Outside contract law the question is open. In Amazon.com Services, LLC v. Perplexity AI, Inc. (opens in a new tab), Amazon sued the maker of an AI assistant that browsed and shopped on Amazon’s site for its users. The Ninth Circuit held that on the record before it, the user accessed Amazon’s computers, and the assistant was “a tool, not a person” under the federal computer-fraud statute (page 15). The court said it was setting no general rules for AI agents, and it left open whether the assistant’s maker could avoid liability in tort (page 17). For now, an agent’s act can run down the chain to the user who sent it.

Florida has no statute that assigns responsibility for a chatbot’s harms, and the activity asks you to draft one. Every liability rule leans one way. Accelerationism lets the technology spread and sorts out harms as they come; decelerationism puts precautions first. Whichever way you lean, the draft has to survive this reading’s defenses. Section 230 bars state laws inconsistent with it (subsection (e)(3)), and a rule about what a chatbot may say will meet the First Amendment. California’s companion-chatbot law, SB 243, and Florida’s SB 482, which passed the Senate and died, show two ways legislators have tried to write around both.

Start the activity with the Defense Tester, then build your first chain from the Florida complaint’s parties and three counts. Put the developer, the deployer, and the user in the rows, and make every defense cite its source.