The FTC found DoNotPay never tested its robot lawyer against a human one

The world's first robot lawyer

For most of its life as a consumer product, DoNotPay sold itself as “the world's first robot lawyer.”¹ The pitch was specific. The company said its artificial intelligence “operated like a human lawyer,” drafting demand letters, small-claims complaints, and contracts, applying the relevant law and steering around complications like statutes of limitations and jurisdiction.¹ Its chief executive, Joshua Browder, described the goal as being “the general counsel for the consumer” and, more expansively, replacing “the $200-billion-dollar legal industry with artificial intelligence.”¹

The people who bought this were ordinary consumers and small businesses. A General Membership ran $36 every two months; a Small Business Protection Plan ran $49.99 a month.¹ For that, a subscriber could ask the service to produce a legal document and get back something that read like one. Whether it read like one and whether it held up as one are different questions, and the distance between them is where the FTC's case lives.

The courtroom stunt DoNotPay called off

The most vivid version of the promise never reached the Federal Trade Commission's file. In January 2023, DoNotPay announced that it would put its AI inside a real courtroom. A defendant fighting a traffic ticket would wear an earpiece; the AI would listen to the proceedings through a phone and feed the defendant arguments to repeat, verbatim, as the hearing went on.² The company offered $1 million to anyone with a suitable case willing to let it try.³

Browder called the plan off within weeks. State Bar prosecutors had warned him that arguing a case this way was the unauthorized practice of law, and he wrote that it “seems likely they will put me in jail for 6 months if I follow through with bringing a robot lawyer into a physical courtroom.”² None of this became part of the FTC's case. The episode is documented in contemporaneous reporting and Browder's own statements, not the Commission's findings, and it is worth holding apart from them. What it shows is what DoNotPay was willing to promise: a lawyer's work in a live courtroom, announced before anyone had shown the AI could do it.

What the FTC found and ordered

In September 2024, the FTC named DoNotPay in a five-case sweep it called Operation AI Comply.⁴ The complaint did not try to prove that any particular document DoNotPay generated was legally wrong. Its charge was narrower, and for that reason harder to answer: the company had made claims about the service that it had no evidence to support.

The service was built on a natural-language-processing model, chatbot software, and a connection to OpenAI's ChatGPT, none of which, the Commission noted, had been trained on a comprehensive and current body of federal and state law.¹ No attorney checked the output. The two counts were both about advertising: that the claim the AI “operates like a human lawyer” was false or unsubstantiated, and that similar claims about a website-diagnostic feature and other membership features were too.¹ One detail shows how far the marketing had drifted from its evidence. DoNotPay's homepage quoted “The Los Angeles Times” praising the robot lawyer as doing work “astonishingly similar” to a human lawyer's. The line was real, but it came from a high-schooler's op-ed on the paper's user-submitted platform, not its newsroom.¹

The case settled. The finalized Decision and Order, issued January 14, 2025, required DoNotPay to pay $193,000.⁵ The order calls this monetary relief, not a fine: the money goes to a Commission-administered fund for consumer redress, and the deception counts here carry no civil penalty.⁵ DoNotPay had described its target as a $200-billion legal industry; the payment that closed the case was $193,000. The order also bars the company from advertising that its service performs like a real lawyer unless it holds “competent and reliable evidence,” meaning tests or studies by qualified professionals. And it requires DoNotPay to email everyone who newly subscribed between January 1, 2021 and December 31, 2023 a notice that its advertising was challenged.⁵ The full record sits on the Commission's docket for the matter: the complaint, the consent agreement, and the final order.⁶ The FTC finalized the order on a 5–0 vote.⁷

DoNotPay settled without admitting or denying the allegations.⁵ Its public response kept to that line. The company said it was “pleased to have worked constructively with the FTC to settle this case and fully resolve these issues, without admitting liability,” and that the complaint “relates to the usage of a few hundred customers some years ago (out of millions of people), with services that have long been discontinued.”⁸

Why a fluent legal document is not a correct one

The underlying failure is not special to DoNotPay. A language model produces text that fits the pattern of the document it was asked for. A demand letter it writes will have the shape, the tone, and the vocabulary of a demand letter. Whether the law inside it is stated correctly, whether a rule it cites exists, whether that rule applies to this person's situation: the model checks none of it. A false statement of law comes out the same way a true one does. It is the same failure that shows up wherever a model writes: fluent, confident text is no guarantee the content is right.

The Commission did not rule that DoNotPay's documents were bad. It found something prior: the company had never checked.¹ A subscriber reading a generated demand letter could see that it was fluent. Whether it would hold up the way a lawyer's letter would was a separate fact about the service, and by the FTC's account no one at DoNotPay had ever put that question to a test. The fluent letter was the thing everyone could see. Whether the service worked like a lawyer was the thing underneath it, which the fluency could neither establish nor rule out.

The useful question is what would count as evidence that a service performs like a lawyer. It would look something like testing the output against work by qualified attorneys, across the kinds of matters the service handles. That is close to the standard the order later imposed: “competent and reliable evidence,” meaning tests or studies by qualified professionals.⁵ The finding was that DoNotPay had gathered none. When AI-drafted legal work does reach a court unchecked, the cost is on the record: two New York lawyers filed a brief citing six cases that did not exist, invented by ChatGPT and never verified. DoNotPay sold the step before that one, the confident document produced for a fee, with no lawyer between the model and the customer.

Where the untested claim is sold now

DoNotPay was one of five cases in Operation AI Comply.⁴ The others were not about law. The FTC named schemes selling AI-driven e-commerce “business opportunities” (Ascend Ecom, Ecommerce Empire Builders, and FBA Machine), and Rytr, an AI writing tool the Commission said could churn out fake reviews.⁴ What they shared was a sales move: a capability asserted, with no evidence it worked as sold. The technology varied; the claim did not. The FTC's chair put the shared line plainly: “there is no AI exemption from the laws on the books.”⁴

The same move is on ordinary product pages now. A tool is “AI-powered.” A feature “uses AI” to do something a professional used to do. Some of those claims rest on real testing and some rest on nothing, and the buyer usually cannot tell which from the page. DoNotPay is a clean example of AI-washing: a service sold on testing that no one performed.