The useful version of ChatGPT for lawyers is a first-draft tool: client letters, website FAQs, intake scripts and marketing copy that a person checks. Used the wrong way, it can get a lawyer sanctioned or put a client’s details into a model’s training data. The task, the plan and the person who checks the output make the difference.
Below, we sort law firm tasks by risk, show which ChatGPT and Claude plans train on your chats (and the setting that stops it) and read three sanctions cases from the court records, then give prompts and a one-page policy. We don’t sell a legal AI tool, so none of it is a product pitch.
This post covers using ChatGPT inside your firm. Advertising on ChatGPT (our October 2026 test) and getting it to recommend your firm (our GEO guide) are separate topics, and our guide to AI for law firms covers the wider practice. This is general information for law firm owners, not legal advice.
Key Takeaways
- OpenAI says that on ChatGPT Free, Plus and Pro “we may use your content to train our models” unless you opt out, but that it doesn’t use Business, Enterprise, Edu or API content by default (OpenAI).
- In Mata v. Avianca, a federal judge imposed a $5,000 penalty on two lawyers and their firm for filing fake opinions ChatGPT invented (S.D.N.Y.).
- ABA Formal Opinion 512 requires a client’s informed consent before information about their matter goes into a self-learning AI tool; engagement letter boilerplate is not enough (ABA).
On this page: Good uses · Sanctions cases · Chatbots · Plans and settings · Opinion 512 · Prompts · Policy starter · FAQ · Sources
ChatGPT works best on drafts a person checks first
ChatGPT is OpenAI’s chat assistant. ABA Formal Opinion 512 describes tools like it as “prediction tools that generate a statistically probable output when prompted” and warns that some are prone to “hallucinations,” “providing ostensibly plausible responses that have no basis in fact or reality” (ABA). That makes it useful for structure, tone and plain English, and unreliable as a source of facts or case law.
The safe use cases share three traits: ChatGPT drafts from material you give it, the draft holds no client details, and a person checks it before anyone relies on it.
The risk ratings below are our judgment, not a rule.
| Task | Who usually does it | Risk | What the reviewer checks |
|---|---|---|---|
| Client letter and email templates (status updates, what to expect at a hearing) | Lawyer, paralegal | Low, with placeholders instead of client facts | Deadlines, tone, no predictions about the outcome |
| Website FAQs and practice area pages | Marketing, lawyer | Low | Each legal statement against the statute or court page; your bar’s advertising rules |
| Intake scripts and follow-up texts | Intake team | Low | No legal advice in the script; the conflict check stays in |
| Plain-English explanation of a statute or court process | Lawyer, paralegal | Low, if you paste the official text | The draft matches the text you pasted |
| Legal research starting point: issues to check, search terms | Lawyer | Medium | Every authority found in a real research tool and read |
| Contract review, or summaries of a client’s records or a deposition | Lawyer, paralegal | Medium to high | Business plan only, consent under Opinion 512, spot-check against the source |
| Legal documents filed with a court, or any case citation | Lawyer | High | The signing lawyer reads every authority cited |
Two hypothetical examples show the gap. A paralegal’s template email explaining an arraignment, with blanks for the client’s name and date, is low risk once a lawyer approves it, and it gets reused every week. An associate’s request for “cases holding that” a claim survives a statute of limitations problem returns leads at best; the next section shows what happens when one goes into a brief unread.
Intake teams get the most from templates. Our law firm intake scripts guide covers what a script needs, and ChatGPT can turn that structure into wording for your practice areas. For marketing work such as blog drafts, ads and review replies, see AI tools for law firm marketing.
Courts sanction lawyers who file what ChatGPT invents
ChatGPT can produce case names, quotations and citations that look real and do not exist, and courts hold the lawyer responsible for filing them. In Mata v. Avianca, U.S. District Judge Castel wrote that “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings” (S.D.N.Y.).
These are not rare. A public database kept by researcher Damien Charlotin listed 2,202 legal decisions involving AI-hallucinated material on October 21, 2026, 1,501 of them in the U.S. (AI Hallucination Cases). Three from the court records, oldest first:
| Case | Court and date | What went wrong | What the court ordered |
|---|---|---|---|
| Mata v. Avianca, Inc. | U.S. District Court, S.D.N.Y., June 22, 2023 | Lawyers submitted “non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT,” then stood by them after the court questioned them | $5,000 penalty, jointly on two lawyers and their firm; letters to the client and to each judge named as an author of a fake opinion |
| Park v. Kim | U.S. Court of Appeals, Second Circuit, January 30, 2024 | Counsel used ChatGPT to find precedent and “did not read or otherwise confirm” the non-existent state case she cited | Referral to the court’s Grievance Panel; a copy of the decision to the client |
| Noland v. Land of the Free, L.P. | California Court of Appeal, September 12, 2025 | 21 of 23 case quotations in the opening brief were fabricated; counsel had “enhanced” his drafts with ChatGPT, run them through other AI tools and filed them without reading them | $10,000 sanction; a copy to the client; the court ordered its clerk to send the opinion to the State Bar |
What checks would have caught each mistake?
Each case points to a different check:
- The signer reads the authorities. In Mata, the lawyer who signed the filing reviewed it for style. “I was basically looking for a flow,” he told the court, and he did not review the judicial authorities it cited.
- Reading is the rule. Under Rule 11, the Second Circuit said, attorneys must “read, and thereby confirm the existence and validity of, the legal authorities on which they rely.”
- Another AI tool is not a check. In Noland, counsel relied on ChatGPT, Claude, Gemini and Grok. Our take: a model asked whether a case exists can vouch for one it made up. The court published its opinion as a warning that no filing “should contain any citations… that the attorney responsible for submitting the pleading has not personally read and verified.”
Legal research AI narrows the problem without removing it. A Stanford and Yale study found that AI research tools from LexisNexis and Thomson Reuters made things up less often than a general chatbot (GPT-4), yet still “each hallucinate between 17% and 33% of the time” (Journal of Empirical Legal Studies). The check is the same for every tool: open each authority in a trusted source, confirm each quote word for word, and confirm the case says what your brief says.
A public chatbot should collect facts, not give legal advice
A website chatbot should collect facts and book consultations, not tell a stranger whether they have a case. OpenAI’s usage policies bar “provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional” (OpenAI). Some state bars add their own guidance: Florida’s advisory Ethics Opinion 24-1 says a lawyer “must inform prospective clients that they are communicating with an AI program” (The Florida Bar). Our AI receptionist guide covers the setup.
Consumer ChatGPT and Claude plans can train on your chats
Whether ChatGPT or Claude learns from what your team types depends on the plan. OpenAI may train on chats from its individual plans (Free, Plus, Pro) unless you opt out, but not on Business, Enterprise, Edu or API content by default (OpenAI). Anthropic uses consumer Claude chats for training if you allow it or a chat is flagged for safety review (Anthropic), and not Claude for Work or API content by default (Anthropic).
The plans side by side, as each company describes them on October 21, 2026:
| Plan | Trains on your chats? | How to keep chats out of training | Watch for |
|---|---|---|---|
| ChatGPT Free, Plus, Pro | Yes, unless you opt out | Turn off “Improve the model for everyone” in Settings > Data controls, or choose “Do not train on my content” in OpenAI’s Privacy Portal | A thumbs up or down can send the whole conversation for training, even after you opt out |
| ChatGPT Business, Enterprise, Edu | Not by default | Nothing to switch on; the feedback buttons don’t permit training | Business workspace admins “can view, access, export, and delete end user conversations” |
| OpenAI API | Not by default | Leave data sharing off at the organization level | If owners turn sharing on, members can share feedback and Playground chats |
| Claude Free, Pro, Max | Only if you allow it, or a chat is flagged for safety review | Turn off “Help Improve our AI models” under Settings > Privacy; incognito chats are not used | A thumbs up or down stores the whole conversation for up to 5 years and may be used for training; if you allow training, de-identified data can be kept up to 5 years |
| Claude for Work (Team, Enterprise), Claude API | Not by default | Owners can turn off the feedback button (“Rate chats”) | A thumbs up or down stores the whole conversation for up to 5 years |
What do the training settings not cover?
Three details trip people up:
- Opting out covers new chats. “After you opt out, we won’t use your new conversations to improve our models,” OpenAI says. Anthropic says data already in a training run, or in a trained model, stays there (Anthropic).
- Temporary chats still sit on a server. A temporary chat isn’t used for training, but “OpenAI may keep a copy of a temporary chat for up to 30 days for safety purposes” (OpenAI).
- Business plans keep data too. On ChatGPT Business, deleted or unsaved conversations “are removed from our systems within 30 days,” with exceptions for legal requirements and harm prevention. OpenAI offers a Data Processing Addendum for Business, Enterprise and API customers (OpenAI).
Our take: put work that touches a client matter on a business plan your firm controls, and turn training off on any personal account used for firm work. Client names stay out of personal accounts entirely.
ABA Formal Opinion 512 puts verification, confidentiality and billing on the lawyer
The American Bar Association’s ethics committee issued Formal Opinion 512 on July 29, 2024. It applies the existing Model Rules to generative AI (“GAI”) tools like ChatGPT. For the security side, see our post on AI agents and law firm ethics risks. These parts shape daily use (ABA):
- Verification. Relying on a GAI tool’s output without “an appropriate degree of independent verification or review” could violate the competence duty in Rule 1.1.
- Confidentiality. Self-learning tools could reveal one user’s input to others. So “a client’s informed consent is required prior to inputting information relating to the representation into such a GAI tool.” Adding “general, boiler-plate provisions to engagement letters” is “not sufficient.” Idea generation that uses no client information needs no consent. Lawyers should also read each tool’s “Terms of Use, privacy policy, and related contractual terms and policies.”
- Former and prospective clients. Rules 1.9(c) and 1.18(b) “require lawyers to extend similar protections to former and prospective clients’ information,” so what a caller tells your intake team counts too.
- Supervision. “Managerial lawyers must establish clear policies regarding the law firm’s permissible use of GAI,” and supervisors must make reasonable efforts to ensure that lawyers and nonlawyers “comply with their professional obligations when using GAI tools.”
- Billing. On an hourly matter, a lawyer who spends 15 minutes entering information to get a draft may bill those 15 minutes plus review time (ABA). A lawyer “may not charge a client to learn about how to use a GAI tool” used regularly for clients.
Our reading: a plan that doesn’t train on your inputs lowers the leak risk the consent rule targets. It doesn’t remove the duty to make reasonable efforts to protect client information. Your state’s rules decide the details, so check your bar’s guidance.
Do courts require you to disclose ChatGPT use?
Some do. In the Northern District of Texas, a brief prepared using generative AI “must disclose this fact on the first page under the heading ‘Use of Generative Artificial Intelligence’” under Local Civil Rule 7.2(f), effective September 3, 2024 (N.D. Tex.). Check your court’s local rules and your judge’s standing orders before filing.
Six ChatGPT prompts for lawyers that need no client data
We wrote these six prompts for common law firm jobs. Each works from text you supply and needs no client details. Replace the bracketed parts.
1. Client update template
You are a paralegal at a [practice area] firm in [state]. Write an email template a lawyer can send after a client’s first court date. Use placeholders such as [CLIENT FIRST NAME], [WHAT HAPPENED] and [NEXT DATE]. Plain English, under 200 words. Make no predictions about the outcome. End with how to reach our office.
2. Website FAQ from the statute
Using only the statute text I paste below, draft six questions and answers for a [practice area] firm’s website in [state]. Keep each answer under 60 words. Do not cite cases. Flag any sentence that goes beyond the pasted text. [Paste the official statute text.]
3. Intake script
Write a phone script for the intake team at a [practice area] firm. Start with a greeting and a conflict check that collects the other parties’ names before any details of the matter. Add questions to see whether the caller fits our practice, a way to answer “how much does it cost?” without quoting a fee, and a step to book a consultation. No legal advice. Mark each line a lawyer must approve.
4. Research starting point
I’m researching [issue, described without names or identifying facts] under [state] law. List the legal questions I need to answer and the search terms to use in Westlaw or Lexis. Do not name or cite any case or statute. Tell me which parts depend on facts I haven’t given you.
5. Plain-English explainer
Explain the court process in the text below to someone with no legal background, in under 250 words, for a client handout. Use only the pasted text. If the text doesn’t answer something a client is likely to ask, list it as a question for the lawyer instead of guessing. [Paste the court’s own page or rule.]
6. Blog outline
Outline a blog post for a [practice area] firm in [city, state] that answers “[a question clients ask].” Give the headings, the official source each section needs (statute, court page, agency) and three questions readers will ask next. Don’t write the post and don’t invent statistics.
The same habits work in any prompt: give ChatGPT a role, paste the text it should use, set the format and length, say what it must not do, and ask it to flag what it isn’t sure of. Break big jobs into smaller requests, and follow up when the first answer misses.
A one-page ChatGPT policy your firm can adopt
Opinion 512 expects managing lawyers to set clear AI policies. This one-page starter is not legal advice, so have your ethics counsel or your bar’s ethics hotline check it against your state’s rules.
- Approved tools only. Firm work goes into the AI tools and plans the firm approves. Anyone using a personal account turns model training off first and keeps client information out.
- What stays out. Client and prospect names, case numbers, medical and financial records, and facts that identify a matter stay out, unless the tool is approved for them and the client has given any consent Opinion 512 requires.
- Every authority gets read. No case, statute or quotation from an AI tool goes into a filing, letter or memo until a lawyer has opened it in a trusted source and read it. Asking another AI tool doesn’t count.
- A lawyer reviews legal content. Anything that states the law or advises a client gets a lawyer’s review before it goes out. Staff may send templates a lawyer has approved.
- No AI legal advice to the public. Website chat and AI receptionists collect information and book consultations. They don’t answer legal questions.
- Court rules first. Before filing, check local rules and the judge’s standing orders for AI disclosure.
- Bill the time worked. Bill actual time, including review, and not the time spent learning a tool the firm uses regularly.
- Marketing gets the same check. Ads, blog posts and FAQs follow your state’s advertising rules, and a lawyer approves every legal statement.
- One owner, two reviews a year. A named person keeps the approved-tool list current, trains new hires on this policy and rechecks plans and privacy settings every six months, since they change.
If you want a second set of eyes on where AI fits in your intake and marketing, a Law Firm Growth Evaluation is where we start.
Frequently asked questions about ChatGPT for lawyers
Is it illegal for a lawyer to use ChatGPT?
No. ABA Formal Opinion 512 doesn’t ban it; it applies existing duties such as competence, confidentiality and supervision. The sanctions cases above punished lawyers for filing output they had not read, not for using the tool.
What is the best AI for lawyers?
Our take: it depends on the job. A general assistant such as ChatGPT or Claude on a business plan handles drafts built from text you supply. For research, a legal research tool makes fewer things up than a general chatbot, per the Stanford and Yale study above, but a person still reads every result.
Is Claude or ChatGPT better for lawyers?
Our take: for the tasks on this page, both write usable first drafts, and both can make things up. Choose on the plan and data terms your firm can accept; both companies say their business plans don’t train on your inputs by default. Then test both on three tasks your team does every week.
Do I have to tell clients or the court that I used ChatGPT?
Sometimes. Opinion 512 says “the facts of each case will determine whether Model Rule 1.4 requires” telling clients, and lawyers must disclose their practices if a client asks. Some courts, such as the Northern District of Texas, require a disclosure on briefs.
Can paralegals and intake staff use ChatGPT?
Yes, with supervision. Under Opinion 512, supervising lawyers must make reasonable efforts to see that nonlawyer staff meet their professional obligations when using AI tools. Give staff approved tools, lawyer-approved templates and a written policy. Intake staff collect facts and book consultations; they don’t answer legal questions.
Use ChatGPT for first drafts and keep a lawyer on what counts
ChatGPT earns its place on templates, FAQs, intake scripts and marketing drafts built from text you supply. Keep client work on a plan that doesn’t train on it, and have a lawyer read every authority before anything is filed or sent. We hold this blog to a similar rule: Claude drafted this post, and a separate check opened every linked source before publication.
Sources
- OpenAI Help Center, “How your data is used to improve model performance,” https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance, checked October 21, 2026.
- Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions (S.D.N.Y. June 22, 2023), ECF 54, copy on CourtListener, https://storage.courtlistener.com/recap/gov.uscourts.nysd.575368/gov.uscourts.nysd.575368.54.0_3.pdf, checked October 21, 2026.
- American Bar Association, “Formal Opinion 512: Generative Artificial Intelligence Tools” (July 29, 2024), https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf, checked October 21, 2026.
- Damien Charlotin, “AI Hallucination Cases” database, https://www.damiencharlotin.com/hallucinations/, checked October 21, 2026.
- U.S. Court of Appeals for the Second Circuit, Park v. Kim, No. 22-2057 (January 30, 2024), copy on CourtListener, https://storage.courtlistener.com/pdf/2024/01/30/park_v._kim.pdf, checked October 21, 2026.
- California Court of Appeal, Second Appellate District, Noland v. Land of the Free, L.P., No. B331918 (September 12, 2025), https://www.courts.ca.gov/opinions/archive/B331918.PDF, checked October 21, 2026.
- Magesh, Surani, Dahl, Suzgun, Manning and Ho, “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools,” Journal of Empirical Legal Studies (2025), https://dho.stanford.edu/wp-content/uploads/Legal_RAG_Hallucinations.pdf, checked October 21, 2026.
- OpenAI, “Usage policies” (effective October 29, 2025), https://openai.com/policies/usage-policies/, checked October 21, 2026.
- The Florida Bar, “Ethics Opinion 24-1” (January 19, 2024), https://www.floridabar.org/etopinions/opinion-24-1/, checked October 21, 2026.
- Anthropic Privacy Center, “Is my data used for model training?” (consumer products), https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training, checked October 21, 2026.
- Anthropic Privacy Center, “Is my data used for model training?” (commercial products), https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training, checked October 21, 2026.
- OpenAI, “Enterprise privacy at OpenAI” (updated January 8, 2026), https://openai.com/enterprise-privacy/, checked October 21, 2026.
- Anthropic Privacy Center, “How do I change my model improvement privacy settings?,” https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings, checked October 21, 2026.
- Anthropic Privacy Center, “How long do you store my data?,” https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data, checked October 21, 2026.
- OpenAI Help Center, “Temporary chat in ChatGPT,” https://help.openai.com/en/articles/8914046-temporary-chat-in-chatgpt, checked October 21, 2026.
- U.S. District Court for the Northern District of Texas, Special Order 2-98, 2024 Local Rules Amendments (effective September 3, 2024), https://www.txnd.uscourts.gov/sites/default/files/orders/2-98.pdf, checked October 21, 2026.
Header image: AI-generated image.
See what turns your marketing
into signed cases.
Your firm doesn't need more random marketing. It needs a growth system that shows what turns into signed cases.