ChatGPT for Lawyers: What It Can and Can't Do (2026)

ChatGPT for lawyers comes down to one line: it's a good first drafter and a bad lawyer. If it's already open in another tab, the useful question isn't whether you're allowed to use it. It's which jobs on your desk it does well, which ones it will quietly ruin, and where the ethics rules draw a line you can't argue across.

This page is written for the attorney, not for their client. It isn't legal or ethics advice, and your state bar's guidance governs over anything here. One disclosure up front: I build DokuTrak, a document request tool, which is why task 11 on the list below is the one I know best. It's also the only task on the list ChatGPT can't touch.

Key Takeaways

• Good fit: client emails, plain-English explanations, marketing copy, intake FAQ text, brainstorming, and rough first drafts, with no client-identifying details in the prompt.
• Wrong tool: finding case law. US courts have sanctioned attorneys in order after order since 2023 for filing citations a model invented.
• ABA Formal Opinion 512 (July 29, 2024) asks for the client's informed consent before client information goes into a self-learning AI tool, and says a boilerplate waiver doesn't meet that bar.
• One job it can't do at all: ask your client for their documents, and keep asking until they arrive.

Can lawyers use ChatGPT?

Lawyers can use ChatGPT for low-risk, non-authoritative work: drafting, summarizing, explaining, brainstorming. No ethics rule bans the tool itself. What the rules require is a working understanding of its limits. They also require the client's informed consent before client information goes into a self-learning tool, and independent verification of anything the tool states as law.

The profession has largely answered the permission question in practice. The American Bar Association's 2024 Legal Technology Survey Report found 30% of responding attorneys using AI-based tools, up from 11% the year before.¹ Published adoption numbers swing wildly between surveys because they poll different populations, so the number is a signal, not a benchmark.

The interesting question is fit. Not "may I," but "on which task, and what's better when it isn't this."

11 tasks lawyers reach for ChatGPT on, and whether it's the right tool

Here is the whole page in one table. Three verdicts: Good fit means a general assistant is the right tool and a legal product adds little. Only with care means usable, with one specific guardrail. Wrong tool means don't, and here's what to use instead. One task at the bottom gets a fourth verdict of its own, because the tool can't perform it at all.

Client-facing writing (tasks 1-4)

This is where a general assistant earns its keep: the work is prose and the stakes are low. A status-letter draft, a plain-English explanation of what a motion to dismiss does, a practice-page rewrite, the FAQ a prospect reads at 11 pm. All drafting, and drafting is what these models are genuinely good at. Two constraints travel with every one. Keep the client's name and facts out of the prompt. And anything the model writes for your website is your advertisement, under your name, with your state bar's advertising rules applying exactly as if you had typed it.

On task 1, if the firm already pays Microsoft, the drafting help is already where the document is: Microsoft 365 Copilot is listed at $18.00 per user per month billed yearly (retrieved 2026-09-20).

Task 4 splits in a way worth naming. ChatGPT writes the FAQ. It can't be put on your phone line to talk to a lead who calls after hours. That's a different product: Smith.ai lists a free tier of 25 calls a month, then $3.00 per call, with a first paid plan at $150/month (retrieved 2026-09-20).

Thinking and drafting (tasks 5-6)

Brainstorming is a good fit with one rule: no output leaves the brainstorm without being run down to a real source. The failure mode isn't the idea. It's the citation that gets attached to it on the way out.

First drafts of routine documents are riskier. A general model writes generic prose against no precedent of yours. Once you draft contracts regularly, a Word add-in that redlines inside the document against your own clause library does more: Spellbook publishes no list price and says pricing is set by the number of team members on a license (retrieved 2026-09-20). For anything case-critical, and demand letters in particular, a rough draft from a general model is the wrong starting point.

Documents and the record (tasks 7 and 9)

Summarizing is where the confidentiality rule actually bites, because the document is client information; the section below on Opinion 512 covers what's expected of you before that paste happens. There's also a structural limit worth naming: you get one summary and no view of which passages the model weighted, so an uneven read of a 300-page transcript looks exactly like a careful one. Products built for this show their work. CoCounsel Legal says it reasons from Westlaw and Practical Law and links every source in an answer so you can check it first (retrieved 2026-09-20).

Task 9 is the row that moved. ChatGPT Record, in the macOS desktop app, captures microphone and system audio and returns notes with timestamped references; it reached Plus users in July 2025, caps a session at 120 minutes, and its notes and transcripts can be referenced in later chats and may be used to improve OpenAI's models.² So it can sit in your meeting now, on one operating system, and that last clause is the confidentiality problem in a sentence. It also isn't a transcript formatted for legal use: Rev lists human transcription at $1.99 per minute and legal-formatted transcripts at $2.25 to $2.50 per page, and Otter.ai lists its Pro plan at $8.33 per user per month billed annually (both retrieved 2026-09-20). Add consent-to-record rules that vary by state, and this is a careful row, not a free one.

The back office (task 10)

Time entries are a good mechanical fit and an ethical trap; Opinion 512 is blunt about billing, and that's covered below. On the mechanics, retyping narratives is the wrong problem to solve: Smokeball sells AutoTime as passive time and activity tracking, and MyCase lists $50 per user per month billed annually with AI features branded 8am IQ, including a writing assistant (both retrieved 2026-09-20).

Where ChatGPT fails: the citations

There is no prompt that fixes legal research in a general model. It has no citator, no authoritative case database, and no step that automatically checks a citation it just generated. Web search can confirm a case exists if you ask; nothing makes that happen on its own, and nothing tells you the case is still good law or says what your draft claims it says. The tool isn't lying. It's completing a pattern, and a plausible-looking citation is an extremely easy pattern to complete.

Courts have been writing this down for three years.

Three things to take from that table. The sanction in every one of these orders is for filing it, not for using it: the failure is skipped verification, and the courts say so in those words. The line doesn't depend on which tool was used, since Wadsworth involved a firm's own in-house system, not ChatGPT. And the trend runs the wrong way, with later orders, larger amounts, and repeat conduct treated as repeat conduct.

The scale is no longer anecdotal. A database maintained by researcher Damien Charlotin, tracking decisions where a court addressed apparently AI-generated content, listed 2,044 cases when I read it on 2026-09-20, with 1,397 of them in the United States.⁸

Switching to a legal research product reduces this problem. It doesn't remove it. In a 2024 study, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Stanford RegLab and HAI researchers found that Lexis+ AI produced incorrect information more than 17% of the time and Westlaw's AI-Assisted Research hallucinated more than 34% of the time.⁹ Those products have changed since, and I'm not presenting 2024 numbers as current performance. The conclusion to draw isn't buy the other tool. It's check the citation, which is why the one task-8 product I'd recommend without hesitation is a checker rather than a generator: Clearbrief tests whether the cases a brief cites actually say what the brief claims, at $300/month per user on its Solo plan (retrieved 2026-09-20). If you want the research itself grounded, note that LexisNexis replaced Lexis+ AI with Lexis+ with Protégé in February 2026, built around its Shepard's citation verification.¹⁰

What you can and cannot put in it

Start with the part every listicle gets wrong: a boilerplate waiver in your engagement letter is not informed consent. ABA Formal Opinion 512, issued July 29, 2024, asks that consent explain why the tool is used, the specific risk, the categories of client information that may be disclosed, and how later users of the tool might access it. A generic clause doesn't carry that, though the engagement agreement is still a sensible place for a specific disclosure.¹¹

Four more points from the same opinion, anchored to the tasks above:¹¹

• Competence. You don't have to become an expert in the tool, but you do need a working understanding of what it can and can't do. How much independent verification a task needs depends on the task: document review demands more scrutiny than brainstorming. (Tasks 6, 7, 8.)
• Confidentiality. Before client information goes into a self-learning tool, evaluate the risk of unauthorized disclosure. Self-learning tools are the specific concern, because input can be used to develop the tool further. (Task 7.)
• Supervision. Managing and supervising lawyers are expected to set a clear firm policy on permitted AI use and to train people on it.
• Fees. You may not bill a client for your own time learning the tool. Under an hourly fee, only time actually worked may be billed, so hours the tool saved you cannot be charged as if you'd worked them. (Task 10.)

On plans: data-handling defaults differ between consumer and business tiers, and that difference is real. Anthropic's pricing page states "No model training on your content by default" for Claude Team, listed at $20 per seat per month billed annually (Anthropic, retrieved 2026-09-20); OpenAI publishes its own terms for its business plans. Read the page for the plan you're actually on, not a summary of it in an article.

Then hold two ideas apart, because every competitor page blurs them. A paid plan changes the training question. It does not answer the consent question. Opinion 512's consent requirement is about your client, not about a vendor's retention policy. I also found no vendor statement addressing attorney-client privilege either way, so I'm not going to tell you what happens to privilege. That one goes to your bar.

ChatGPT prompts for lawyers that are safe to run

Six prompts, one per good-fit task. Every one of them is written to be safe to run with no client-identifying information in it, which is the point of the first line in each. Structure is always the same: role, constraint, input, output format.

1. Routine client letter. "You are drafting for a US attorney. Here is a fact pattern with all names and identifying details removed: [situation]. Write a 200-word client update letter in plain English, no legal conclusions, neutral tone. End with a list of what I should insert." → Check that it made no promise about outcome or timing.

2. Plain-English explanation. "Explain [legal concept] to a non-lawyer at an 8th-grade reading level in under 150 words. Describe the law in general. Do not apply it to any specific situation." → Check the description of the law, then say out loud that it's general, not advice about their matter.

3. Marketing copy. "Write three versions of a 120-word practice-area page introduction for [area of law] in [state]. No superlatives, no claims about results, no statistics." → Check it against your state's attorney advertising rules before it goes live.

4. Intake FAQ. "Write eight questions a prospective client asks before hiring a [practice area] attorney, with 60-word answers. No fee figures. No outcome predictions." → Check that no answer accidentally gives legal advice to a non-client.

5. Brainstorming. "Here is an anonymized fact pattern: [facts]. List the strongest arguments on each side, then the three weakest points in my position. Do not cite cases." → Check nothing leaves this list without being run down to a real source.

6. First draft of a routine document. "Draft a [document type] between two unnamed parties with these terms: [terms]. Use neutral drafting. Flag every place where a term is missing or ambiguous with [BRACKETS]." → Check every bracket, every defined term, and every cross-reference yourself.

There is no research prompt on this list, and that's deliberate. A prompt for finding case law would undercut the entire page. There is no prompt that fixes it.

The one thing ChatGPT cannot do at all

ChatGPT has no outbox, no client list, and no idea who still owes you a bank statement. Not a prompt problem. A chat window can't send a request to your client, can't track what arrived against what you asked for, and can't follow up on Tuesday. That's the structural limit, and it sits on the single task most likely to be blocking your matter right now.

If your firm already runs practice-management software, check whether it has a client portal: MyCase includes one from its entry plan up (retrieved 2026-09-20). For plenty of firms that's enough. The gap opens with the clients who never log in. They keep sending documents by email and text. Nothing tells you page 2 of the lease is missing or the statement is for the wrong year until someone opens it. And the follow-up is still typed by hand.

Run your own arithmetic on that last part. A firm opening 8 new matters a month, with 3 documents per matter that arrive late, each needing 2 follow-ups of about 5 minutes to find the thread, write it and note it: 8 × 3 × 2 × 5 = 240 minutes. About 4 hours a month on those assumptions, which scheduled reminders can take over. Your numbers are different; the shape usually isn't.

That's the job DokuTrak does. One itemized request per matter, which you can start from the free law firm client intake checklist. One secure upload link, with no account for the client. Scheduled reminders written as file status rather than nagging: Your file: 4 of 6 documents received. And an AI first-pass check that flags wrong, unreadable or expired files, which you then accept or send back.

To be exact about what that check is: it confirms a file is the right kind of document and readable. It doesn't assess legal sufficiency and it doesn't give legal advice. Nothing here does. For the email-and-reminders half of the same problem, I wrote how to stop chasing clients for documents and a set of document request email templates you can use today, with or without software.

When to stop using ChatGPT and buy something

Three triggers, in the order they usually arrive. You draft the same document type every week: buy the drafting tool that works in Word against your own precedent. You cite authority in filings: buy the citator-backed research product, the citation checker, or both, and treat it as the cost of Rule 11 rather than a productivity upgrade. You spend more time chasing documents than reading them: not a ChatGPT problem at all, and no prompt will touch it.

Everything else on the good-fit list, ChatGPT already does. Keep the client's name out of it, verify anything it calls law, and have the consent conversation before the first document goes in.

Send your first request free. First client request free, no card. Then a 14-day trial, no upfront charge. Solo is $79/month, Team $199, Agency $449.

Footnotes

1. LawNext (LawSites), "ABA Tech Survey Finds Growing Adoption of AI in Legal Practice, with Efficiency Gains as Primary Driver," March 7, 2025, reporting the American Bar Association's 2024 Legal Technology Survey Report (512 attorneys in private practice). Retrieved 2026-09-20. https://www.lawnext.com/2025/03/aba-tech-survey-finds-growing-adoption-of-ai-in-legal-practice-with-efficiency-gains-as-primary-driver.html ↩

2. Juli Clover, "ChatGPT Plus Gets Record Mode on Mac for Meeting Transcription," MacRumors, July 17, 2025. Retrieved 2026-09-20. https://www.macrumors.com/2025/07/17/chatgpt-plus-gets-record-mode-macos/ ↩

3. Eugene Volokh, "Sanctions Issued in Case Where Lawyers Cited ChatGPT-Hallucinated Precedents," Reason / The Volokh Conspiracy, June 22, 2023. Retrieved 2026-09-20. https://reason.com/volokh/2023/06/22/sanctions-issued-in-case-where-lawyers-cited-chatgpt-hallucinated-precedents/ ↩

4. Federal Public Defender news, "Second Circuit Seeks to Sanction Attorney for Fake Citation in Latest AI Blunder," on Park v. Kim, No. 22-2057 (2d Cir. Jan. 30, 2024). Retrieved 2026-09-20. https://www.fd.org/news/second-circuit-seeks-sanction-attorney-fake-citation-latest-ai-blunder ↩

5. LawNext (LawSites), "Federal Judge Sanctions Morgan & Morgan Attorneys for AI-Generated Fake Cases in Court Filing," February 2025, on Wadsworth v. Walmart Inc. (D. Wyo., Feb. 24, 2025). Retrieved 2026-09-20. https://www.lawnext.com/2025/02/federal-judge-sanctions-morgan-morgan-attorneys-for-ai-generated-fake-cases-in-court-filing.html ↩

6. Eugene Volokh, "$6K Sanctions for Apparent AI Hallucinations in Coomer v. Lindell," Reason / The Volokh Conspiracy, July 7, 2025. Retrieved 2026-09-20. https://reason.com/volokh/2025/07/07/6k-sanctions-for-apparent-ai-hallucinations-in-coomer-v-lindell-my-pillow-election-related-libel-suit/ ↩

7. Eugene Volokh, "$5K Sanctions for Repeated Mis-Citation in Coomer v. Lindell," Reason / The Volokh Conspiracy, May 9, 2026. Retrieved 2026-09-20. https://reason.com/volokh/2026/05/09/5k-sanctions-for-repeated-mis-citation-in-coomer-v-lindell-my-pillow-election-related-libel-suit/ ↩

8. Damien Charlotin, "AI Hallucination Cases" database, page stating "Last updated: September 19, 2026." Retrieved 2026-09-20. https://www.damiencharlotin.com/hallucinations/ ↩

9. Stanford HAI, "AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries," May 23, 2024, on Magesh, Surani, Dahl, Suzgun, Manning and Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Retrieved 2026-09-20. https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries ↩

10. LawNext (LawSites), "LexisNexis Launches Lexis+ with Protégé, Replacing Lexis+ AI with an End-to-End Workflow Platform," February 24, 2026. Retrieved 2026-09-20. https://www.lawnext.com/2026/02/lexisnexis-launches-lexis-with-protege-replacing-lexis-ai-with-an-end-to-end-workflow-platform.html ↩

11. NCBE, The Bar Examiner, "Generative Artificial Intelligence Tools," Fall 2024, summarizing ABA Formal Opinion 512 (July 29, 2024). Paraphrased, not quoted from the ABA opinion. Retrieved 2026-09-20. https://thebarexaminer.ncbex.org/article/fall-2024/generative-artificial-intelligence-tools/ ↩ ↩²