When I teach the ethics hour on generative AI, I start by asking the room a show of hands: who has pasted something from a client file into ChatGPT? The hands go up slowly, then all at once. Then I ask who read the data settings first. Most of the hands go back down.
That gap is the whole subject of this article. GPT for lawyers is not a question of whether the technology works. It plainly does, for a good range of tasks. The question is whether the lawyer using it can meet the duties that already bind them: competence, confidentiality, candor to the tribunal, supervision and reasonable fees. None of those rules were rewritten for AI. They simply apply, and courts have spent the past three years showing what happens when lawyers forget that.
TL;DR: Lawyers can use ChatGPT and other general GPTs ethically, but only for work where a mistake is cheap to catch, and only with confidentiality settings checked and every authority confirmed by hand. ABA Formal Opinion 512 (July 2024) sets the framework. For research you intend to cite, a purpose-built legal tool such as LegesGPT, which answers with verified citations and source links, closes most of the gap a general chatbot leaves open.
Can lawyers use ChatGPT at all?
Yes. No ABA rule and, as far as I am aware, no U.S. state bar prohibits a lawyer from using a general-purpose GPT. What the bar has done is explain how existing obligations attach to the tool.
The anchor document is ABA Formal Opinion 512, issued by the Standing Committee on Ethics and Professional Responsibility on July 29, 2024. It is short enough to read over lunch, and I recommend you do. It works through six Model Rules:
- Rule 1.1 (competence): you need a reasonable understanding of what the tool can and cannot do, including that it can produce false output.
- Rule 1.6 (confidentiality): before you put client information into a tool that may retain or learn from it, you generally need the client's informed consent, and the opinion says boilerplate in an engagement letter is not enough.
- Rule 1.4 (communication): in some circumstances you must tell the client you are using AI, for instance when its output will inform a significant decision or the client asks.
- Rules 3.1, 3.3 and 8.4(c) (meritorious claims and candor): you are responsible for every assertion of law and fact you file, whoever or whatever drafted it.
- Rules 5.1 and 5.3 (supervision): partners and supervising lawyers need policies and training for lawyers and staff who use these tools.
- Rule 1.5 (fees): if you bill hourly, you bill for the time actually spent, not the time the work would have taken without AI, and you generally cannot bill clients for learning a tool you will use across your practice.
Many states have issued their own guidance since, and several track the ABA analysis closely. Check your own jurisdiction, because a few go further on disclosure, and an increasing number of individual judges have standing orders requiring certification of AI use in filings.
What a general GPT actually does well in a law practice
I am not here to scare anyone off the tool. Used on the right tasks, ChatGPT earns its subscription in the first week. The common thread in the tasks below is that the lawyer can judge the output on sight, without needing to check it against an outside source.
Rewriting and tightening prose. Give it a paragraph from your own draft and ask for a version half the length, or a version a twelfth-grader could follow. You know what you meant, so you will spot any drift in meaning immediately.
Structuring a document you already understand. Asking for an outline of a client advisory on a new state privacy law, or the section headings for a motion in limine, saves the blank-page problem. You fill in the substance.
Generating issue lists to check against your own. Before a client meeting on a commercial lease renewal, ask what issues a tenant typically raises. If the list contains something you had not thought of, go research it properly. If it misses something, you will notice.
Explaining unfamiliar territory at a high level. A litigator who suddenly needs the basics of ERISA preemption can get a plain-language orientation in five minutes. That orientation is the start of research, never the end of it.
Administrative writing. Engagement letter cover emails, scheduling correspondence, internal training materials, marketing copy for the firm website. Low stakes, easy to review.
Summarizing non-confidential material. A published opinion, a public regulation, a long article from a bar journal. Nothing in the input belongs to a client.
Notice what is missing from that list: finding case law, confirming a statute is still current, and analyzing a client's documents. Those are the three places where general GPTs cause the most trouble.
Where it goes wrong: the hallucination problem
A large language model predicts plausible text. A case citation is a very predictable string of text: a party name, a "v.", another party name, a reporter volume, a page and a year. The model can produce a perfectly formatted citation for a case that has never existed, and it will often summarize the holding with complete confidence.
The case every CLE instructor uses is Mata v. Avianca, Inc. in the Southern District of New York. Counsel used ChatGPT for research on a personal-injury claim against an airline and filed a brief citing six decisions that did not exist. When opposing counsel and the court could not find them, the lawyers did not immediately come clean. On June 22, 2023, Judge P. Kevin Castel imposed a $5,000 penalty under Rule 11, finding subjective bad faith, and required the lawyers to send letters to the judges who had been falsely named as authors of the fake opinions. The docket is on CourtListener if you want to read the order in full.
It would be comforting to treat Mata as an early-adopter mistake. The record says otherwise:
- In Park v. Kim, 91 F.4th 610 (2d Cir. 2024), the Second Circuit referred a lawyer to its Grievance Panel after her reply brief cited a nonexistent case she had generated with ChatGPT, and ordered her to give her client a copy of the decision.
- In Johnson v. Dunn (N.D. Ala. July 2025), Judge Anna Manasco disqualified three lawyers at a large regional firm from the case after discovery motions cited authority fabricated by ChatGPT. The firm had a written AI policy. It did not stop the filing.
- Researcher Damien Charlotin maintains a database of AI hallucination cases in courts worldwide. As of early October 2026 it lists more than 2,100 decisions, around 1,470 of them in the United States, and over 850 involving lawyers rather than self-represented parties.
The lesson I draw for students is not "AI lies." It is that the duty under Rule 11 and Rule 3.3 is personal and non-delegable. In Johnson, the lawyers who signed had not drafted the offending sections themselves. The court held them responsible anyway.
Hallucination is also subtler than invented cases. Real cases get cited for propositions they do not support. Quotations get slightly reworded so they no longer match the opinion. Overruled or superseded authority is presented as good law. A general GPT has no built-in way to tell you which of its statements came from a source and which it composed.
The confidentiality question nobody reads the settings for
Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure of client information. With a general GPT, "reasonable efforts" starts with knowing which version you are using.
As OpenAI describes its own data controls, conversations on consumer ChatGPT plans (Free, Plus and Pro) may be used to improve its models unless the user turns off the "Improve the model for everyone" setting under Data Controls. Content from ChatGPT Business, Enterprise, Edu and the API is not used for training by default. Even with training switched off, a consumer conversation is still stored on OpenAI's servers under its retention terms, and clicking the thumbs-up or thumbs-down on a response can send that conversation for review.
Put plainly: a solo practitioner on a personal Plus account who pastes a draft settlement agreement into the chat window, without changing a setting, has made a disclosure decision on the client's behalf. Opinion 512 treats tools that may learn from inputs as the case where informed client consent is most clearly required.
Three practical controls get most firms to a defensible position:
- Use a business-tier account, or a legal tool with a written commitment not to train on your data, for anything that touches a matter.
- Strip identifying details before you paste. "A Delaware LLC supplier, a 36-month term, a $2.4 million minimum purchase commitment" carries the analysis without naming the client.
- Write the rule down. A one-page firm policy that says which tools are approved for which data is the kind of measure Rules 5.1 and 5.3 contemplate.
Custom GPTs for law: a feature with an expiry date
Many lawyers built a "custom GPT for law" over the past two years: a ChatGPT configuration loaded with firm instructions, a style guide and a handful of reference documents. Some were useful. If you rely on one, note that OpenAI has announced it is retiring custom GPTs across ChatGPT plans in favor of plugins, which bundle reusable instructions (now called skills) with connected apps. Creation of new custom GPTs is being phased out this autumn, and the scheduled retirement date is December 11, 2026, with a deferral to February 2027 for some Enterprise workspaces. Instructions and knowledge files can be migrated, but conversation history and custom actions do not carry over, and OpenAI itself warns that a migrated plugin may answer differently.
From an ethics standpoint, the more important point is that a custom GPT never changed the underlying model's relationship with the truth. Uploading twenty of your favorite opinions as "knowledge" did not stop it inventing a twenty-first. Lawyers sometimes treated a firm-branded GPT as though it were vetted. It was the same general model wearing a firm letterhead.
OpenAI's own move into legal-specific products confirms the point. In September 2026 it announced Astra for Law, a configuration of its GPT-6 Astra model paired with a legal search index, offered by invitation to selected large firms through its Trusted Access program. The company evidently agrees that a general chatbot, on its own, is not the right instrument for legal research.
General GPT vs. legal-specific GPT
The distinction that matters is not the brand of model underneath. It is whether the tool's answers are tied to retrievable legal sources and whether its data terms were written with privileged material in mind.
| Question to ask | General GPT (ChatGPT and similar) | Legal-specific GPT tool |
|---|---|---|
| Where do legal answers come from? | Model training data, plus open web search when enabled | Retrieval from case law, statutes and regulations, with the source attached |
| Can you click through to the authority? | Sometimes a web link; often nothing | Expected: each citation links to the underlying document |
| Risk of fabricated case citations | Well documented in sanctions decisions | Reduced, though not zero; still read what you cite |
| Training on your inputs | On by default for consumer plans unless switched off | Usually contractually excluded; confirm in the terms |
| Built for privileged material? | Business tiers have stronger terms; consumer plans do not | Generally designed around confidentiality from the start |
| Document review workflow | Paste text and ask; no structured risk flags | Upload a contract, get clause-level issues and suggested edits |
| Cost for a solo lawyer | Roughly $20 a month for a consumer plan | Ranges from about $20 a month to enterprise contracts |
| Best use | Prose, outlines, brainstorming, non-confidential summaries | Research you will cite, client document review, drafting from legal templates |
Neither column removes the lawyer from the loop. Opinion 512 is explicit that the degree of independent checking depends on the tool and the task, but there is no tool for which the answer is "none."
Safe prompting habits I teach in CLE sessions
These are habits rather than magic words. They reduce risk; they do not transfer responsibility.
Ask for structure, not authority. "List the elements a plaintiff must prove for tortious interference with a contract under Texas law" invites a useful checklist. "Give me three Texas cases holding X" invites invention. Get the elements from the model if you like, then find the cases yourself in a citator-backed source.
Tell it the facts are hypothetical and anonymized, and make sure they are. Replace names, dates and amounts with placeholders before you paste. If the anonymized version would still let someone identify the client, it is not anonymized.
Demand that it flag uncertainty. Adding "if you are not sure a rule or case exists, say so rather than guessing" does not eliminate fabrication, but it does make the model more likely to hedge, and hedges are easier to spot than confident errors.
Paste the source in, then ask about it. A model asked to summarize a statute you have supplied is on much firmer ground than a model asked to recall that statute from memory. Even then, check that quotations in the summary match the text word for word.
Keep the chat out of the final product. Treat anything the model gives you as a colleague's rough note. Rewrite it into your own work, and do not paste it straight into a filing or a client email.
Log what you used. A short note in the matter file recording which tool you used and what you checked costs thirty seconds and answers the question a judge or a client may later ask.
If you want a broader look at how practitioners are fitting these habits into day-to-day matters, the walkthrough on using AI across legal work covers workflows beyond the ethics angle.
Supervision: the duty that catches firms, not just individuals
Rules 5.1 and 5.3 require lawyers with managerial authority to make reasonable efforts to ensure that the firm has measures giving reasonable assurance that everyone, lawyers and nonlawyers alike, complies with the Rules. Opinion 512 applies that directly to generative AI.
Johnson v. Dunn is the cautionary example here too. The firm had a policy. The court still found that lawyers signed filings without reading the citations. A policy that nobody follows is not a "measure" in any meaningful sense. What a supervising partner can reasonably put in place:
- A list of approved tools, with what data each may receive.
- A rule that every citation in a filing is pulled and read by a human before signature, regardless of who drafted the section.
- Short, repeated training, especially for junior associates and paralegals who are most likely to reach for ChatGPT under deadline.
- A clear path for someone to say "I used AI on this section" without fear, so the reviewing lawyer knows where to look harder.
When a purpose-built legal tool is the better choice
My rule of thumb: once the output will be relied on by a court, a client or an opposing party, use a tool built for that job. Three situations come up repeatedly.
Research you plan to cite. A legal research tool that grounds its answer in actual case law and statutes, and links each citation to the source, changes the checking task from "does this case exist?" to "does this case say what the summary says?" That is a much faster task. Our roundup of AI legal research tools compares the main options on exactly this point.
Client contract review. Pasting a 40-page supply agreement into a chat window gives you a summary. A review tool gives you clause-level risk flags and proposed redlines, under data terms written for confidential material. The comparison of AI contract review tools covers how the leading products handle it.
Drafting from legal forms. Starting from a vetted template and adapting it is safer than asking a general model to invent a contract from scratch, which can omit boilerplate a court will later read literally.
For solo lawyers and small firms, LegesGPT is the tool AIForLegal recommends for this work. Every answer it gives comes with verified citations that open the underlying source; it also lets you search case law and statutes, marks risky clauses in uploaded documents with suggested changes, drafts contracts from more than 100 templates, and includes a Deep Research mode for longer questions plus built-in e-signature. It runs in the browser only (there is no Word add-in or mobile app), and plans start at $19.99 a month for Basic, with Plus at $49.99 a month adding 50 document reviews and Premium at $99.99 a month for unlimited reviews. A 3-day trial costs $1. Citations you can click through still need to be read, but you start from real authority instead of a guess.
If you are weighing several assistants rather than just one, our guide to AI legal assistants for lawyers sets them side by side.
A short self-audit before your next prompt
Before using a general GPT on a matter, I ask attendees to answer five questions honestly:
- Do I know whether this account trains on what I type, and have I checked the setting myself?
- Is there anything in this prompt that could identify my client or the matter?
- Will I be able to tell, without outside research, if the answer is wrong?
- If the answer contains any case, statute or quotation, have I set aside time to pull and read the original?
- If a judge or client asked tomorrow how this work product was prepared, would I be comfortable with the true answer?
If any answer is no, either change the prompt, change the tool, or do the task the conventional way.
The bottom line on GPT for lawyers
General GPTs are good writing and thinking aids, and an ethical lawyer can use them every day. They are poor research tools, unreliable citators, and, on consumer settings, the wrong place for client confidences. The rules governing all of this have not changed since long before ChatGPT launched: know your tools, protect your client's information, stand behind every word you file, and supervise the people and systems working for you.
Match the tool to the stakes. Use ChatGPT for the first draft of an email; use a legal research platform with linked authority for the brief. The lawyers in Mata, Park and Johnson did not lose because they used AI. They lost because they let it sign for them.
Frequently Asked Questions
Can lawyers use ChatGPT ethically?
Yes. No ABA rule bans it, and ABA Formal Opinion 512 from July 2024 confirms lawyers are free to use generative AI as long as they meet their existing duties of competence, confidentiality, communication, candor, supervision and reasonable fees. The practical test is whether you can protect client information and personally confirm every legal authority before relying on it.
What is the biggest risk of using GPT for lawyers?
Fabricated or misstated authority. General chatbots can produce realistic citations to cases that do not exist, quote real opinions inaccurately, or present overruled law as current. Courts have sanctioned, referred for discipline and even disqualified lawyers who filed such citations, including in Mata v. Avianca, Park v. Kim and Johnson v. Dunn.
Does ChatGPT keep or train on what I type?
On consumer plans such as Free, Plus and Pro, conversations can be used to improve OpenAI's models unless you switch off the Improve the model for everyone setting in Data Controls. Business, Enterprise, Edu and API content is not used for training by default. Either way, check the current terms and avoid pasting identifiable client information into a consumer account.
Do I need client consent to use a legal GPT on their matter?
Often, yes. ABA Formal Opinion 512 says informed consent is generally required before entering client confidences into a tool that may retain or learn from them, and that boilerplate language in an engagement letter is not sufficient. Some state bars add their own disclosure expectations, so check your jurisdiction.
Are custom GPTs for law still available in ChatGPT?
OpenAI has announced it is retiring custom GPTs in favor of plugins, with creation of new GPTs being phased out and a scheduled retirement date of December 11, 2026, and later deferrals for some Enterprise workspaces. Instructions and knowledge files can be migrated, but chat history and custom actions do not carry over, so firms should test any replacement before relying on it.
How is a legal-specific GPT different from ChatGPT?
A legal-specific tool retrieves answers from case law, statutes and regulations and links each citation to its source, and it usually comes with data terms written for confidential material. ChatGPT draws mainly on its training data and general web search. Both still need a lawyer to read the authority before it is cited.
What is a safe way for lawyers to prompt ChatGPT?
Ask for structure and explanations rather than case citations, anonymize facts before pasting, supply the source text yourself when you want a summary, ask the model to flag uncertainty, and rewrite the output in your own words instead of pasting it into a filing. Keep a short note in the file of what tool you used and what you checked.
What should a lawyer use instead of ChatGPT for legal research?
For research you intend to cite, use a tool built for legal work that grounds answers in real authority. LegesGPT backs each answer with a citation you can open, lets you look up case law and statutes directly, and flags risks in uploaded documents, with plans from $19.99 a month and a 3-day trial for $1. You should still read each authority before filing.
