AI legal advice chatbots and the unauthorized practice of law

The short version

• An AI chatbot can give general legal information. Applying the law to one person’s own facts, for them, is the practice of law, which US states reserve to licensed lawyers.
• The US Federal Trade Commission (FTC) fined the “robot lawyer” DoNotPay, Inc. US$193,000 for deceptive claims, not for the unauthorized practice of law; a concurring FTC commissioner said the Commission does not enforce state unauthorized-practice law.
• The professional-conduct rule, ABA Model Rule 5.5(a), bars a lawyer from practising law in a jurisdiction in violation of that jurisdiction’s regulation of the profession, or from assisting another in doing so, the clause that can reach a lawyer who fronts an AI tool.
• Courts have treated an automated legal-document service as the practice of law because of the human legal judgment programmed into it, not because it runs online.
• Court-run sandboxes in Utah and non-lawyer-ownership reform in Arizona are testing whether to let AI and software deliver legal help under supervision, to widen access to justice.^{3,5,6,7,8,9,11,18,19,21}

What is the unauthorized practice of law?

The unauthorized practice of law (UPL) is a non-lawyer doing work the law reserves to licensed lawyers. Most US states police it by statute and through the courts’ own authority.^6,8

The line the whole question turns on is legal information against legal advice. General facts about the law and how a process works are information. Applying legal rules to a particular person’s own facts is advice, and giving that advice for a fee is the practice of law. The FTC’s order in the DoNotPay matter lists the lawyer-like work at issue: applying the relevant laws to a subscriber’s particular situation, relying on legal expertise to generate a demand letter or start a small-claims case, and detecting legal violations and advising how to fix them. In In re Reynoso a federal appeal court held that selecting legal exemptions and giving advice went past the clerical help a non-lawyer may provide.^3,11

The practice of law is defined by each jurisdiction, not by any single national rule. California’s statute, Business and Professions Code section 6125, reserves it to active licensees of the State Bar. The common thread across the rules is non-delegation: a named person stays answerable for legal work.^7,8

Can an AI chatbot give legal advice, or is that unauthorized practice?

A chatbot can give legal information lawfully. It moves toward the unauthorized practice of law when it applies the law to a user’s own facts and the user relies on what it says as advice. No single authority settles every product, because the practice of law is defined state by state.^6,8,9

This holds for a general-purpose chatbot such as ChatGPT as much as for an app built for legal work. Asking it to explain a rule, or how a process runs, is legal information. Asking it what to do about your own facts is where it drifts toward advice.^6,8

Whatever a tool does, a lawyer who uses AI stays bound by the existing duties of competence, confidentiality, candour and supervision, the American Bar Association’s ethics committee says.¹²

How courts and lawyers use AI, and how the law regulates AI itself, is covered in AI in law. What the tools actually do is set out in AI legal tech. Where AI is given a goal and plans, takes several steps and acts through tools, such as running a search or filling a form, rather than only answering a single prompt, an emerging label calls it agentic AI; see agentic AI in dispute resolution.

Why did the FTC act against the “robot lawyer”?

In September 2024 the FTC brought an action against DoNotPay, Inc. as one of five cases in a sweep it called Operation AI Comply. The complaint says the company marketed “the world’s first robot lawyer” but did not test whether the service’s law-related features operated like a human lawyer, and had “not employed attorneys and has not retained attorneys” to check them.^1,2

The order, Docket C-4812, bars the company from representing that its product “operates like a human lawyer”, including “by applying the relevant laws to subscribers’ particular legal and factual situations”, unless it holds “competent and reliable evidence” for the claim. It requires US$193,000 in monetary relief and notice to subscribers who joined between 1 January 2021 and 31 December 2023. The Commission voted 5 to 0 to approve the final order on 16 January 2025 and announced it on 11 February 2025; the order document itself bears an issue date of 14 January 2025.^3,4

This is a deception case under Section 5 of the FTC Act, not an unauthorized-practice case. In a concurring statement filed with the September 2024 complaint, a commissioner said his vote “should not be taken as support for the State Bar of California’s claim that DoNotPay was engaged in the unauthorized practice of law”, and that “[t]he Commission does not enforce state occupational-licensing laws like California’s unauthorized-practice-of-law prohibition”. The FTC record names only DoNotPay, Inc. as the respondent.^3,5

What does the professional-conduct rule say?

ABA Model Rule 5.5(a) says a lawyer shall not practise law in a jurisdiction in violation of that jurisdiction’s regulation of the profession, or assist another in doing so. The “or assist another” clause is the hook that can reach a lawyer who builds or fronts an AI legal tool.⁶

The model rule is not law on its own. It binds a lawyer only once a state’s highest court adopts it. California’s adopted rule 5.5 has the same structure: a lawyer shall not “knowingly assist a person in the unauthorized practice of law in that jurisdiction”, and its comment points to Business and Professions Code section 6125 and following. California’s rule took effect on 1 November 2018.^7,8

A comment to the rule lets a lawyer delegate work to non-lawyer assistance under supervision while keeping responsibility for it. Read that way, a tool a lawyer supervises and stands behind is different from an unsupervised product sold straight to the public as a lawyer substitute. That reading is interpretation, not the words of the rule.⁶

Bar ethics bodies have already applied existing duties to generative AI and chatbots, as persuasive rather than binding guidance: the American Bar Association’s Formal Opinion 512, the Florida Bar’s Opinion 24-1, the New York State Bar Association’s task force report and the State Bar of California’s practical guidance.^12,13,14,15

Have courts found software or a non-lawyer service practising law?

Yes, in several cases, though the outcome turns on state law. What decided each one was how much human legal judgment the service carried, not that people reached it over the internet.^9,10,11

Read each posture with care, because they do not carry equal weight. The firmest is TIKD, a final 2021 decision of the Supreme Court of Florida, which held that disclosing non-lawyer status does not cure the problem. Reynoso, from 2007, is an appellate holding about bankruptcy-petition software. Janson weighs least: a 2011 district-court summary-judgment ruling, so it persuades rather than binds.^9,10,11

A different failure mode is a filing that contains fabricated, AI-invented citations, known as hallucinations. There the court sanctions the person who signed the filing; that record is collected in the AI hallucination sanctions record.

Are regulators testing a different approach?

Yes. Rather than only policing unauthorized practice, a few US courts have changed their rules to let non-lawyer and technology providers deliver legal services under supervision. Utah runs a time-bound sandbox pilot to test whether this widens access without more harm to consumers, and Arizona has made a permanent change by licensing non-lawyer-owned firms.^18,21

Utah’s legal regulatory sandbox, set up by the Utah Supreme Court under Standing Order No. 15, lists “technology-based services such as AI” among the provider types it can authorise. It is a time-bound pilot: it is closed to new applications, the alternative-business-structure portion closed at the end of 2024, and the whole sandbox is set to sunset on 14 August 2027, with recommendations pending. That is the position as of 10 November 2026.^17,18,19

Arizona has licensed alternative business structures, which let non-lawyers own or run a law firm, since 1 January 2021. In England and Wales, the Solicitors Regulation Authority authorised an AI-based firm for small-claims debt recovery while naming regulated solicitors as accountable and stating that the system “is not autonomous”. The newest US statutes that reach AI used in legal work are set out in US state AI laws: Colorado and Texas.^16,20,21

How does this play out in dispute resolution and access to justice?

The same line runs through dispute resolution, and the access-to-justice case cuts across borders even though unauthorized practice is mainly a US doctrine. A tool can give the parties information, but applying the law to one party’s own facts, or letting AI decide or mediate the outcome, moves into reserved work and puts a neutral’s independence in question. A neutral is the arbitrator, mediator or ombudsman who decides a case or helps the parties settle it.^6,8,22,23

When a chatbot gives a customer bad advice, the business that deployed it is liable, not the bot. In Moffatt v. Air Canada, 2024 BCCRT 149, a British Columbia tribunal held an airline responsible for what its chatbot told a customer and rejected the airline’s argument that the chatbot was responsible for itself.²⁴

That access-to-justice concern has drawn official attention outside the United States. The OECD recommends responsible digital transformation of justice with trustworthy digital tools. The Legal Services Board in England and Wales reports that AI legal tools show real promise for access, but that consumers using them lack basic protections.^22,23

The overview and lead for this section is AI in dispute resolution. AI court decisions inside arbitration are collected in AI arbitration case law. Institutional AI guidance is covered in the SVAMC AI guidelines and tracked in the AI rules tracker. How the EU regulates AI, by contrast with the US approach, is in the EU AI Act in dispute resolution. AI-made or doctored exhibits, called deepfakes, convincing AI-made images, audio or video of something that did not happen, are covered in AI-generated evidence and deepfakes and, inside arbitration, in AI document review.

What does this mean for anyone using AI in a dispute?

In our view, the unauthorized-practice rules and the FTC’s action run on different tracks but point the same way. A chatbot can hand someone legal information. The moment it applies the law to that person’s own facts and they rely on it, the work starts to look like the practice of law, which the states reserve to licensed lawyers. The FTC did not call DoNotPay unauthorized practice. It called the “robot lawyer” claim deceptive and left licensing to the state bars. Read together, the safest position is that the duties already on the books govern AI use. A named person stays responsible for the work. The consumer is told plainly what the tool can and cannot do, and the output is checked before anyone acts on it. The sandboxes are testing whether to widen that line for access to justice, not to erase it.^{5,7,9,10,11,12,18,23}