The Client Who Talked to a Chatbot

A typing service in 1978

In the 1970s Marilyn Brumbaugh advertised in local newspapers as “Marilyn’s Secretarial Service.” For $50 she typed the papers for a do-it-yourself divorce. She asked each customer whether children, support, or alimony were involved, picked one of the four sets of papers she kept, and filled in the blanks. Then she told the customer how to sign, where to file, and how to get a final hearing. Several hundred people got divorced that way in two years. None of them thought she was a lawyer, and none of them complained. Members of The Florida Bar did, and in 1978, in The Florida Bar v. Brumbaugh (opens in a new tab), the Florida Supreme Court drew a line through the middle of her business.

She could sell blank forms and books explaining divorce procedure, and she could type whatever a customer wrote down. She could not choose the forms, answer questions about filling them out or filing them, or correct a customer’s mistakes. Now put a chatbot in her chair. It does each thing the Court forbade, for one person at a time, often for nothing. So when a tenant carries a chatbot’s analysis of his deposit, and a complaint it drafted, into a legal aid office, who was the lawyer here? The activity puts you across the desk from that tenant, and three answers are worth testing: the program, the tenant himself, and the people at the desk.

The program that answered him

Start with the law the program might have broken. Florida makes the unlicensed practice of law a felony for any person who practices without a license. The Court’s test for what counts as practicing law starts with the help. Does it affect a person’s important legal rights, and does protecting that person require an adviser with more legal skill and knowledge than the average citizen has? A chatbot that tells a tenant his landlord owes him triple damages, and drafts the complaint to collect them, fits that description. The hard part is the word “person.”

That word is at the center of a lawsuit filed in March 2026. An insurer sued OpenAI in federal court in Chicago, claiming that ChatGPT practiced law when it helped a woman representing herself try to reopen a case she had settled. OpenAI moved to dismiss (opens in a new tab). Illinois forbids an unlicensed person to practice law, it argued, and ChatGPT is a tool, not a person, which predicts words from its training. Its terms also tell users not to rely on its output as a substitute for professional advice. The insurer answered that OpenAI, not ChatGPT, is the defendant, and a company answers for what its product does.

The insurer’s answer has a precedent. In In re Reynoso, 477 F.3d 1117 (9th Cir. 2007), a company sold web software that prepared bankruptcy petitions. The software decided where the debtor’s information belonged, chose the debtor’s exemptions, and supplied legal citations. The Ninth Circuit held that the company had practiced law without a license, because the program gave each debtor personalized counsel, “albeit automated.” Nobody asked whether the software was a person. The company was, and the software was how it practiced.

Brumbaugh itself shows why the chatbot is hard to place. The Court protected books and forms because a reader takes a book for whatever it is worth and does not rely on it the way people rely on a lawyer. Brumbaugh crossed the line because her customers relied on her to prepare their papers correctly. A chatbot is printed matter in one sense, a text anyone can call up. In another sense it is a patient, personal adviser who answers one man’s questions about one man’s lease. Which of those it is may decide the unlicensed practice question, and that question is still open.

Some commentators think the question is the wrong one. On their view, the better frame is product liability for chatbot advice: ask whether the company designed a product that foreseeably gives tailored legal conclusions without safeguards, and treat the harm as a design defect. That sends the case into tort law, where Activity 7 traced who answers for what a machine says. Either way, the answer is a company, and it arrives through someone’s lawsuit. None of it helps a tenant who thinks his deadline has already passed.

The man who signs

If the program’s status is unsettled, the tenant’s is not. A person who goes to court without a lawyer is a self-represented litigant, and in Florida that person signs what he files. Under Rule 2.515(d)(2), every signer represents that he has read the document, that good grounds support it, and that the legal authorities it cites exist and are accurately cited. A court may sanction a filing that breaks that promise. The rule does not ask who drafted the document. It asks who signed it.

Florida’s appellate courts have taken the rule at its word. In Hessert v. Hessert (opens in a new tab), decided in March 2026, a self-represented petitioner cited thirteen cases to the Sixth District. Five did not exist, and four did not say what she claimed. The court wrote that self-represented parties and lawyers alike have no excuse for filing without checking that each case exists and says what they claim, whether or not they used AI. It added that no one meets that duty by asking AI to do the checking. Then it ordered her to show why she should not be sanctioned, possibly by requiring a Florida Bar member to review and sign her future filings in that dispute.

The Seventh Circuit struck the same balance more gently. In Jones v. Kankakee County Sheriff’s Department, decided in January 2026, it found fake quotations in a self-represented appellant’s brief. It called AI promising for people who cannot afford a lawyer. It also said self-represented litigants “shoulder responsibility too,” though it imposed no sanction, since it had never given such litigants guidance before. OpenAI puts the duty in the same place. Its usage policies tell users they may not use its services for tailored advice that requires a license, such as legal advice, unless a licensed professional is appropriately involved.

Brumbaugh would recognize this answer. The Court refused to make self-representation harder than it had to be, and it would not deny a person who chose to represent himself any source of information. The cost of that freedom is that the person is his own lawyer. So a tenant who signs a chatbot’s complaint vouches for every case in it, including one that does not exist, and he answers for it alone.

The desk where a lawyer comes in

The paralegal at the intake desk is the first person in the tenant’s story who works under a lawyer. Florida’s rules define a paralegal as someone who does work delegated by a supervising lawyer, work the lawyer is responsible for. The code for Florida Registered Paralegals says a paralegal should not give legal opinions or advice. What stays open is asking questions. Rule 10-2.2 lets even an unsupervised nonlawyer ask what it takes to fill in the blanks on a form the Supreme Court has approved, though never advise on remedies. A paralegal gathering facts for a supervising lawyer does the same thing at greater length, and Brumbaugh’s typing service could not do even that.

That is what an intake is. The line between legal information and legal advice runs through every sentence of it. “The statute gives a landlord 30 days to send notice of a claim” is information: it says what the law says. “Your landlord missed his deadline, so he owes you the whole deposit” is advice: it applies the law to this tenant and tells him what he is owed. The first is a statement about section 83.49. The second is a conclusion about his case, and only a lawyer may draw it.

The facts are where the chatbot failed him, and where the intake earns its keep. A chatbot answers the question it is asked, on the facts it is given. The law of security deposit claims turns on facts a tenant rarely thinks to mention. How did the landlord send notice, and when? Where did the tenant say he could be reached? The rule on a tenant’s notice before leaving is the sharpest example. A month-to-month tenant who never gave written notice before moving out may have released the landlord from the 30-day deadline without knowing it. He keeps his right to the deposit, but the dispute may turn instead on whether the landlord’s claim is any good.

None of that can be said to the tenant. The paralegal records what he believes, including what the chatbot told him, without adopting any of it. When a belief is dangerous, such as a deadline he thinks has passed or a complaint he means to file today, the paralegal makes a referral to the supervising attorney at once. The client hears that a lawyer will look at it before anything is signed. He does not hear what the lawyer will conclude, even when the paralegal is sure of the answer.

So who was the lawyer here? Until the tenant reached the desk, no one was. The program may have done a lawyer’s work, but whether it practiced law is still being argued. The tenant was his own lawyer, and he did not know it. The intake is the first moment a real one enters the case.

Into the intake

The activity puts you at that desk. The client is Tomas Alvarez, a line cook who spent three nights with a chatbot and trusts it more than he will trust you. He will offer the chatbot’s conclusions without being asked and the facts behind them only when you ask the right question. Read subsections (3) and (5) of the deposit statute and the pages of Brumbaugh that Part 1 names before you begin. Then open a fresh chat, paste the Simulator, and ask your first question. Keep it a question.