Three real cases that define AI lawyers in court

The blunt answer to whether an ai lawyer in court can appear as counsel in a U.S. proceeding is no. The more useful answer starts in a New York courtroom, where Jerome Dewald tried to present an AI-generated avatar as his legal representative in an employment dispute. The judge did not treat the moment as a charming test of courtroom technology. The court shut it down, and the reported rebuke was direct: “I don’t appreciate being misled.”[1]

That is the line most of the public discussion manages to blur. A court appearance is not merely the delivery of words. It is an act performed under court authority by someone entitled to appear, subject to professional duties, sanctions, privilege rules, candor obligations, and responsibility for what is said. Software can generate an argument. It cannot, under current U.S. law, stand before the court as counsel.

The past two years have produced three useful boundary markers: Dewald’s stopped avatar in New York, DoNotPay’s abandoned “robot lawyer” courtroom plan, and Garfield AI’s English trial win. They are often discussed as signs that AI representation is almost here. Read closely, they say something narrower and more practical: courts and regulators are drawing a hard distinction between AI-assisted legal work and AI appearing as the legal representative.

What the phrase “AI lawyer” gets wrong

“AI lawyer” is a loose phrase doing too much work. It can mean a human lawyer who uses AI, a legal chatbot sold to consumers, a document automation system, a regulated legal services entity using AI internally, or an avatar speaking in a courtroom. Those are not variations on the same legal status.

For U.S. purposes, the central question is not whether an output sounds legal. It is who is practicing law, who is authorized to do so, who supervises the work, who signs or files the papers, who speaks to the tribunal, and who bears professional responsibility if the work is wrong or misleading. For a fuller definitional split between human AI-using lawyers and AI systems, see what an AI attorney actually means.

That distinction matters because unauthorized-practice-of-law rules are not triggered only by bad technology. They are triggered by the performance of legal work by someone or something not authorized to perform it. Every state has UPL restrictions, Model Rule 5.5 prohibits lawyers from assisting unauthorized practice, and courts retain authority to control who appears before them. Those doctrines leave little room for an AI system to present itself as counsel in a U.S. courtroom.

Dewald: the courtroom appearance problem

Dewald’s case is the cleanest example because the attempted use occurred at the point of maximum court control. A reported 74-year-old New York plaintiff in an employment dispute tried to present an AI-generated avatar as his representative at Westchester Supreme Court. The presiding judge halted the attempt after concluding the court had been misled about what was being presented.[1]

The important fact is not that the representative was synthetic. It is that the court was asked to accept a courtroom speaker without the ordinary authority and accountability of an admitted advocate. Once the judge viewed the presentation as misleading, the technology issue became a candor issue. That is usually the faster route to trouble.

Dewald does not appear, on the available materials, to be a final merits ruling on whether every possible AI courtroom use is unauthorized practice. It is still a strong procedural marker. A judge need not wait for a bar prosecution to control who addresses the court, what is disclosed, and whether a party’s presentation is creating confusion about representation.

For lawyers and legal operations teams, that is the first operational lesson: disclosure and role clarity are not decorative. If an AI system prepared a script, summarized exhibits, or helped draft an argument, the person appearing must still be authorized to appear and must own the presentation. If an AI avatar is put forward as the speaker, the question immediately shifts from productivity to authority.

DoNotPay: the “robot lawyer” claim met UPL risk before it reached court

DoNotPay is less useful as courtroom precedent because its proposed “robot lawyer” did not ultimately appear in court. It is more useful as a warning about marketing a tool as a substitute for counsel. State bar prosecutors reportedly threatened the company’s founder with six months’ jail time for unauthorized practice of law, and the courtroom use was abandoned.[2]

The later consumer-protection consequence was separate but related in tone. DoNotPay settled Federal Trade Commission charges for $193,000 after allegedly falsely claiming to be “the world’s first robot lawyer.”[2] The point is not that every consumer legal app is unlawful. The point is that replacing-lawyer language invites regulators to ask whether the product is performing legal representation, inducing false reliance, or crossing a UPL boundary.

There is an access-to-justice argument here that should not be waved away. Many people cannot afford counsel for small disputes, traffic matters, benefits problems, or routine consumer issues. But unmet legal need does not give software a law license, and it does not make a courtroom experiment harmless. A user who relies on overstated claims may lose a deadline, waive an argument, disclose sensitive facts, or misunderstand who owes them professional duties.

DoNotPay therefore clarifies a different boundary from Dewald. Dewald shows a court stopping an attempted AI appearance. DoNotPay shows that prosecutors and regulators may act before any appearance occurs, especially where the public-facing claim sounds like independent legal representation. The UPL trajectory for AI chatbots has its own complications, but the marketing posture matters; tools sold as lawyer replacements attract a different risk profile from tools sold as supervised drafting or intake aids. See also AI chatbot UPL liability cases.

Garfield: the hard case because the AI did real litigation work

Garfield AI is the case worth spending time on because it is not the easy prohibition story. In England, an HR consultant won an unfair dismissal claim at Wandsworth county court using Garfield AI, described as the first known case of an AI law firm winning a trial. Garfield AI had been authorized by the UK Solicitors Regulation Authority in April 2025.[3]

The work division matters. The AI handled the pre-trial work, including witness statements and document bundles, and the reported cost was £400. A human barrister conducted the oral advocacy.[3] That is not the same thing as an AI appearing in court. It is AI-heavy case preparation followed by human courtroom advocacy.

Garfield is significant precisely because the AI’s contribution was not trivial. Witness statements and bundles are not clerical garnish. They shape what the advocate can prove, what the judge sees, and how the case is organized. In many disputes, especially modest-value employment or consumer matters, reducing the cost and friction of that preparation could make the difference between abandoning a claim and presenting it coherently.

But Garfield’s result should not be imported into the United States as if it answers the UPL question. The UK authorization of Garfield AI as an entity is a regulatory fact with no U.S. equivalent in the materials provided. No U.S. state bar has authorized an AI system to practice law. No U.S. jurisdiction has created a parallel pathway under which an AI legal services entity may independently provide representation in court.

That leaves Garfield as the outer edge, not the new baseline. It shows that AI can be deeply involved in litigation without being the courtroom advocate, if the legal system authorizes the entity and a human professional still performs the advocacy role. For U.S. lawyers, the safer comparison is not “AI won a trial.” It is “AI prepared substantial litigation materials, while an authorized human appeared and argued.”

Where the U.S. line currently sits

Under current U.S. law, the line is clearest at the courtroom door. An AI system cannot appear as counsel, examine witnesses, make objections, address the judge as the party’s advocate, or independently represent a litigant. Those acts belong to admitted lawyers, properly authorized advocates, or self-represented parties speaking for themselves.

Behind the scenes, the analysis is more granular. AI may assist with research, chronology building, transcript review, first-pass drafting, deposition preparation, issue spotting, exhibit organization, and bundle assembly when a human lawyer or other authorized professional supervises the work and remains responsible for the final product. That does not make the AI a lawyer. It makes the AI a tool in a workflow governed by professional responsibility rules.

• If AI drafts a motion, a lawyer must review, correct, and sign it before filing.
• If AI summarizes authorities, a lawyer must verify the cases, quotations, procedural posture, and jurisdictional fit.
• If AI prepares witness materials, a lawyer or authorized professional must ensure the statement is accurate, admissible, and properly adopted.
• If AI helps a self-represented litigant, the product must not be marketed or used in a way that suggests an attorney-client relationship where none exists.
• If AI is used in court presentation, the person addressing the tribunal must be authorized and candid about the role of the technology when disclosure is required.

The unresolved area is direct client-facing work that looks less like internal drafting assistance and more like legal judgment delivered to a user. The available materials do not establish that any U.S. jurisdiction has definitively treated an AI system as a permissible nonlawyer assistant under Model Rule 5.3 for independent consumer-facing advice. That uncertainty is not a technicality. It is the place where a product’s design, disclaimers, supervision model, and marketing claims can change the legal analysis.

Disclaimers help, but they do not turn the platform into counsel

The Heppner ruling is a useful supporting note because it reinforces the ordinary status of AI platforms: they are not lawyers and do not create attorney-client relationships merely by producing legal-looking text. In Heppner v. OpenAI, the court noted that Claude AI disclaims providing legal advice, in the course of rejecting a privilege theory over AI-generated documents.[4]

That point cuts both ways. A disclaimer may help defeat a user’s claim that the platform was their lawyer. It may also underscore that the user had no lawyer at all. If the work requires legal judgment, privilege protection, litigation strategy, or court appearance authority, the absence of an attorney-client relationship is not a harmless footnote.

This is why “no attorney-client relationship” language should not be treated as a universal shield. It may reduce one category of confusion while leaving another intact: the user may still believe the system has given legal advice that can safely be followed. For compliance teams, the more important controls are product scope, escalation to counsel, user-facing claims, jurisdictional limits, and documentation of human review.

The practical boundary for litigation teams

The practical boundary is not hard to state, even if particular workflows require care. AI can support litigation work. It cannot be the litigation representative in a U.S. courtroom.

A defensible AI litigation workflow keeps three things visible. First, the AI’s task is bounded: research support, draft generation, document organization, factual comparison, or preparation assistance. Second, an authorized human reviews and adopts the work before it affects a filing, argument, or client decision. Third, the court and client are not misled about who is responsible.

Dewald failed at the point of appearance and candor. DoNotPay drew enforcement attention through the claim that a product could function as a “robot lawyer.” Garfield succeeded, on the reported facts, because the AI-heavy work happened within a different regulatory structure and the oral advocacy remained with a human barrister.

That makes Garfield the most interesting model now available, not because it proves AI lawyers have arrived, but because it strips the phrase down to a workable division of labor. AI can prepare. A regulated professional can appear. The human advocate remains the one standing before the court.

For U.S. practice, that remains the boundary. No U.S. jurisdiction has yet created an equivalent authorization for AI to practice law or appear as counsel. Until one does, the safest description is also the most accurate one: AI-assisted litigation, not an AI lawyer in court.

References

1. AI-generated lawyer sparks courtroom controversy in New York — Fox5NY.
2. The World's First Robot Lawyer Short-Circuited by Prosecutors, Now Faces Class Action Lawsuit — NH Bar Association.
3. HR consultant wins English court case using AI lawyer in apparent legal first — The Guardian.
4. AI Is Not Your Lawyer: Federal Court Rules AI-Generated Documents Are Not Privileged — BakerHostetler.