LexED2 home
AI & the Law: all pages

Who Answers for What a Machine Says and Does?

PLA 2872Activity 7

AI and Free Speech, with Eugene Volokh and Jane Bambauer, Hoover Institution

Why it Matters

Tort law was built for harms with an author: a driver who ran the light, a maker who shipped the cracked ladder, a newspaper that printed the false charge. A chatbot’s words have no single author. A developer trained the model, a deployer built it into a product, and a user typed the prompt, so every question about duty, defect, or publication now has three candidates. When the software acts instead of talking, buying or booking for someone, the law must also decide whose act it was. You trace two real records link by link, from the company that built the machine to the person who used it, and then draft the Florida statute you think the law still lacks.

Current Context

On September 8, 2026, a federal judge sent Florida’s suit against OpenAI and Sam Altman back to state court (opens in a new tab), ruling that its reliance on a federal children’s privacy law did not make the claims federal. The same day, Attorney General James Uthmeier proposed legislation that would make a company criminally liable (opens in a new tab) when an AI system it owns, controls, or distributes takes part in a crime. In Tallahassee, the federal court now holds six civil suits against OpenAI over the April 2025 shooting at Florida State University, the latest removed from state court on September 17. Each complaint on the public docket pleads design defect, failure to warn, and negligence. In the first of them, OpenAI previewed its answers (opens in a new tab) in July: the First Amendment, no duty to monitor what users type, and doubts about foreseeability and other causes.

Key Concepts

Liability Chain

The people and companies between a machine’s output and the harm it caused: the developer who trained the model, the deployer who offered it in a product, and the user who prompted it. A plaintiff needs a theory for each link it sues, and each link has its own defense. One company can hold two links, as when the maker of a model also runs the chatbot built on it.

Design Defect

A flaw in a product’s plan, judged in Florida by whether the product performs as safely as an ordinary consumer would expect or whether its risks outweigh its benefits. The plaintiff need not prove that a safer design existed. A chatbot claim must first show that what failed is a product: courts separate an app’s design, which can be one, from the words and ideas it delivers, which generally are not.

Failure to Warn

A product can be well designed and still defective because it reached the user without adequate warnings or instructions about a risk its maker knew or should have known. The claim asks what the warning said, whether it reached the person at risk, and whether a better one would have changed what happened. For a chatbot, the argument usually turns on the disclaimer: what it said, and whether the person harmed ever saw it.

Duty and Foreseeability

Negligence starts with a duty of care, and in Florida a defendant owes one when its conduct creates a foreseeable zone of risk to others and it is in a position to control that risk. Duty is a threshold question for the judge; whether the conduct foreseeably caused this particular injury is proximate cause, usually a question for the jury. A federal court in Florida has held, at the pleading stage, that releasing a chatbot to the public can create that zone of risk.

Publication and Fault

Florida’s defamation claim has five elements: publication, falsity, fault, actual damages, and a statement that is defamatory. Fault means knowledge or reckless disregard of the falsity when the plaintiff is a public official or public figure, and at least negligence when the plaintiff is a private person. A chatbot’s false statement about a real person tests two of those elements at once: who published the words, and whose state of mind can supply the fault.

Section 230 and the First Amendment

Section 230 of the Communications Decency Act bars treating an online service as the publisher of information provided by another information content provider, so a platform is generally not liable for its users’ posts. The statute defines that provider as anyone responsible, in whole or in part, for creating the information, so a company whose model wrote the words may be its own provider and outside the shield. The First Amendment defense asks a different question, whether a model’s output is protected speech, and courts have not settled it.

Electronic Agent

Florida’s term, in its Uniform Electronic Transaction Act, for a computer program used on its own to start an action or respond to records without review by a person. Under section 668.50(14), a contract can form through an electronic agent even when no person saw what it did. An AI agent, software that shops, books, or files for its user instead of only answering, raises the same question outside contract law, where courts have barely begun: whose act was it?

Accelerationism and Decelerationism

Two stances toward a new technology. An accelerationist favors fast deployment and light regulation, on the view that the benefits of innovation outweigh its harms; a decelerationist favors caution and rules that come before the harm. Every liability rule leans one way or the other, and a drafter should know which way before writing.

Resources

What to Do

A lawsuit over a machine’s harm is a set of choices about whom to name, on what theory, and against which defense. You make those choices link by link on two real records: a Florida complaint against the maker of ChatGPT, and a Delaware ruling on a chatbot’s statements about a real person. Then you turn what you found into a short Florida statute. What you post is the two chains and your statute, with the changes a critic asked for and what you did with each.

Part 1: Read the Two Records

Start with the background reading, then watch the video under the title from 12:13 to about 28:00, where two law professors disagree about how far liability should reach. Then open the Florida complaint in Resources and read the parties at paragraphs 9 to 16. Read three counts after them: Negligence (paragraphs 241 to 249), Strict Liability for a Design Defect (259 to 275), and Strict Liability for Failure to Warn (276 to 288). Each count repeats the long statement of facts before it by reference, and you do not need those facts for this work. Read every sentence of a complaint as an allegation the State must still prove. Then read the Delaware court’s opinion in Starbuck v. Google, pages 1 and 2 and pages 12 to 31. Keep the Garcia order at hand for the product and speech questions, and run the Defense Tester once before you start Part 2.

Part 2: Build Two Liability Chains

Build one table for each record. Its rows are the three links: developer, deployer, and user. Its two columns are the theory that reaches that link, with the paragraph or page that supports it, and the strongest defense that link has, with its source. Where one company holds two links, say so in both rows. Where the record names no one at a link, write who that would be and whether anyone sued them. Under each table, write one sentence naming the link where you think the case will be won or lost, and why. Score your own chains before you open any prompt. Each row earns a point if it names a party from the record, a point if its theory cites a paragraph or page, and a point if its defense cites its source. Then run the Liability Chain Analyzer in a fresh chat for each record, pasting the excerpt you relied on and your chain. It asks questions and fixes nothing. Change a row only where the record supports the change, and record your score again.

Part 3: Take a Stance and Draft

Decide whether you take the accelerationist or the decelerationist stance, and state it in one sentence above your draft. Then draft a Florida statute of about 300 words that fills the gap your chains found. Model its form on California’s companion-chatbot law, SB 243, and on Florida’s SB 482, both in Resources. It has five parts. The title says what the law does. Two definitions let a court apply the law without guessing. One of them says whether the law reaches software that acts, such as an electronic agent under section 668.50, or only software that talks. One duty or liability rule names who owes it and to whom. The enforcer is named with the remedy it can seek, and the last part is an effective date. Score your draft one point for each part it gets right. Then run the Reality Check, which ends with three changes. Make or reject each one, and revise the draft once.

Part 4: Post Your Chains and Statute

Create a single new discussion thread. Post your two chains as tables, each with its one-sentence verdict and both of its scores. Below them, post your stance and your revised statute. End with the Reality Check’s three changes, each followed by a sentence or two on what you did with it and why. Before you post, check each row against its three points and the statute against its five parts.

Include one AI chat link with a 1–2 sentence explanation of what the conversation shows and why you chose to share it.

Suggested AI Prompts

Use these prompts as a starting point, then adjust them to fit your goal. Strong prompting develops through trial, revision, and testing. It’s a foundational skill that grows into more advanced AI work such as context engineering and agent-based workflows.

Defense Tester

Act as a law professor quizzing a first-year student on two defenses that AI companies raise when sued over what their systems say: Section 230 of the Communications Decency Act and the First Amendment. I will paste the text of 47 U.S.C. § 230(c)(1) and (f)(3). Ask me one question at a time. Start with what the statute protects and why, then whether a chatbot’s own output is “information provided by another information content provider,” then how the First Amendment question differs from the Section 230 question. After each answer, tell me what I got right, correct what I got wrong by quoting the statute text I pasted, and then ask the next question. Stop after six questions and name the two points I most need to review. Do not say what a court has held unless you can name the case, and tell me when a question is still open.

A model that asks before it explains makes you do the recalling, which is Socratic elicitation, and one question at a time keeps it from lecturing. Tying each correction to the statute text you pasted keeps the quiz honest on a question the courts are still working out.

Liability Chain Analyzer

Act as a legal analyst who specializes in products liability, negligence, and defamation. I will paste an excerpt from a court record about harm caused by an AI system, followed by my liability chain for it: a table whose rows are the developer, the deployer, and the user, and whose two columns are the theory that reaches each link and that link’s strongest defense, each with the paragraph or page I relied on. Check every row against the excerpt. For each row, tell me whether the paragraph or page I cite supports what I wrote, name any party or defense the excerpt raises that I missed, and ask me one question that would lead me to the fix. Do not rewrite or complete any row, and do not tell me the correct entry. Name any case or statute you rely on precisely, and if you are not sure it exists or says what you think, say so instead of guessing. Wait for me to paste the excerpt and the chain.

Checking each row against the paragraph you cite is source grounding: the model has to read the record you gave it, not its memory of the case. Its admission when it is unsure of an authority tells you which of its points to verify before you change anything.

Reality Check

Act as a skeptical supervising attorney. I am going to share a draft statute on AI liability. Stress-test it: identify unintended consequences, enforcement problems, constitutional concerns such as First Amendment or preemption issues, conflicts with existing tort doctrine, and any gap between the harm the statute targets and the conduct it actually regulates. Be direct about serious flaws and note genuine strengths. Do not rewrite the statute; evaluate it, and end with the three changes you would make first. Confirm that you understand and ask me for the draft.

Asking the model to attack a draft instead of improving it is red-teaming, and it returns weaknesses you can weigh for yourself. The cap of three changes keeps the revision yours, because you decide which to make and say why.

Search LexED2

Type to search every activity, reading, and instruction page. Press Esc to close.