ChatGPT for Lawyers (2026): Safe Uses, Real Risks, and Better Tools

ChatGPT is a useful tool for an in-house lawyer, as long as you treat it as a drafting assistant and not a source of law. It is fast and strong at first drafts, summaries, and rewriting. It is unreliable at citations, jurisdiction-specific rules, and anything you have not fed it yourself. This guide covers the tasks where it earns its keep, the duties that govern its use, the consumer versus business data split that decides whether you can paste a client matter, and what courts now expect when AI touches a filing.

Adoption is already mainstream. In the 8am 2026 Legal Industry Report, 69% of the more than 1,300 legal professionals surveyed between September and October 2025 said they use general-purpose AI tools for work (8am). In the 2025 report, 31% said they personally used generative AI at work (ABA Law Technology Today); the two questions are worded a little differently. The lawyers getting value out of it share one habit. They know which tasks to hand it and which to keep away from it.

Reviewed and corrected on 12 October 2026. OpenAI's plans and terms and court rules change, so check the sources linked below before relying on any of this. This is general information, not legal advice.

TL;DR

• Use ChatGPT for drafting, not for law. First drafts, summaries, rewrites, and brainstorming are where it shines. Finding authority and confirming a holding are where it fails.
• It can fabricate citations that look real. A made-up case looks exactly like a real one, which is why filings with invented authority keep drawing sanctions.
• Consumer and business plans have different data terms. OpenAI's help page says data sharing for training is on by default for Free, Plus and Pro on a personal workspace, with an opt-out. It says it does not train on ChatGPT Business (the plan formerly called Team), Enterprise or Edu inputs and outputs by default (OpenAI).
• ABA Formal Opinion 512 sets the baseline. Read the tool's terms of use and privacy policy, verify its output to the degree the task requires, and get a client's informed consent before inputting information relating to the representation into a self-learning tool.
• Some courts have their own AI rules. For example, the Northern District of Texas requires a brief prepared with generative AI to say so on its first page.
• A research tool that links to the source makes checking cheaper. It does not make checking optional.

What ChatGPT actually is (and is not)

ChatGPT is a large language model that predicts plausible next words. That one fact explains both its talent and its danger.

It is excellent at language: structure, tone, summarizing, rephrasing. It has no internal database of what is true. Asked for a case that supports a point, it can generate a citation-shaped string that looks right. Sometimes that string is a real case. Sometimes it is not, and the model cannot reliably tell the difference.

So the useful mental model is simple. ChatGPT is a very fast junior who writes well, has read a lot, and will not always tell you when it is guessing. You would never file a junior's memo without checking the cites. Same rule here.

Where ChatGPT is genuinely good for lawyers

These are the tasks where a general chatbot saves real time, with the verification burden low because you supply the substance.

• First-draft scaffolding. Hand it a fact pattern and ask for a demand letter, a deposition outline, an internal policy memo, or a client update. You get a coherent structure in seconds that you then edit and fact-check.
• Summarizing text you provide. Paste a long email thread, a deposition transcript, or a contract you already have rights to, and ask for a summary or the key dates. It is working from your text, not its memory.
• Rewriting and tone-shifting. Soften a blunt internal note for a client, tighten a rambling paragraph, or turn bullet points into prose. The law in the document is already yours.
• Brainstorming. Counterarguments, issue-spotting checklists, negotiation angles, questions to ask a vendor. You are using it to think, not to cite.
• Plain-language explanations of a concept you will confirm elsewhere before relying on it.

For a head-to-head on the drafting tier specifically, see our best AI chatbot for legal writing comparison.

Prompts that actually work

Generic prompts get generic output. The good ones do three things: give the model a role, hand it the source material, and constrain the format. Notice that every prompt below supplies the facts and the governing text. None asks ChatGPT to remember the law.

• First draft from your terms: "You are drafting for the licensor. Using the deal terms below, write a mutual NDA with a two-year confidentiality period, a carve-out for independently developed information, and Delaware governing law. Deal terms: [paste]."
• Summarize a document you already hold: "Summarize this master services agreement in under 200 words. List the parties, the term, the auto-renewal mechanics, the liability cap, and any indemnity. Quote the exact clause number for each. Text: [paste]."
• Explain a redline, not invent one: "Here is a vendor's proposed limitation-of-liability clause and my company's standard. Explain in plain English how the vendor's version shifts risk to us and what it removes. Do not rewrite either clause. Clauses: [paste]."
• Tone and audience shift: "Rewrite this internal risk memo as a four-sentence update for a non-legal executive. Keep every factual claim identical. Memo: [paste]."
• Issue-spot as a checklist: "I am negotiating a SaaS reseller agreement. Give me a checklist of the ten clauses most likely to be contested and the business reason each one matters. Do not cite cases or statutes."

The pattern holds across all of them: you own the substance, ChatGPT shapes the language. These prompts assume the pasted text is something you may put into the tool; see the data section below before pasting a client's document. For a wider set organized by task, see our generative AI prompts for in-house lawyers.

Where ChatGPT is bad, and why it matters

The failure modes are not random. They cluster around exactly the work that goes to a court or a client.

Citations and hallucinations

This is the headline risk. A chatbot can produce a case name, a real-sounding court, and a reporter cite in the correct format, all for a case that does not exist. The output looks like competent legal work because the model is built to produce fluent text, not to check facts. ABA Formal Opinion 512 says some generative AI tools are "prone to 'hallucinations,' providing ostensibly plausible responses that have no basis in fact or reality."

Courts have sanctioned lawyers for it.

• Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) (S.D.N.Y. June 22, 2023). Judge P. Kevin Castel found that two lawyers and their firm had submitted non-existent judicial opinions with fake quotes and citations created by ChatGPT, and then stood by them after the court questioned them. He ordered sanctions under Rule 11, or alternatively the court's inherent authority: letters to the plaintiff and to each judge falsely named as an author of a fake opinion, and a $5,000 penalty imposed jointly and severally on the respondents. The same opinion says there is "nothing inherently improper about using a reliable artificial intelligence tool for assistance," but that attorneys have a gatekeeping role to ensure the accuracy of their filings (order, ECF 54).
• Couvrette v. Wisnovsky, No. 1:21-cv-00157-CL (D. Or.). On December 12, 2025, the court struck three summary-judgment briefs that cited fifteen non-existent cases and carried fabricated quotations attributed to eight real authorities. It ordered the pro hac vice lawyer to pay the clerk $15,500 and awarded the defendants their fees, dismissed the plaintiffs' claims with prejudice, and ordered local counsel to show cause (opinion, ECF 215). On March 23, 2026, the court set the fee award at $94,704.38, ordered local counsel to pay 15% of it ($14,205.66) for failing to meaningfully participate in the case, and ordered the pro hac vice lawyer to pay the other 85% ($80,498.72) (opinion, ECF 225). Those figures plus the $15,500 come to $110,204.38. The court's opinion refers to misuse of generative AI and does not say which tool. We break the matter down in the Oregon $110k AI hallucination sanction.

Lawyers keep filing authority they have not checked because the fakes are convincing and the deadline is real. We track the running pattern, and the fix, in AI hallucinations and legal research sanctions.

What courts now require

Court rules on AI differ by court, so check yours before you file.

• Northern District of Texas. In May 2023 Judge Brantley Starr issued a standing order requiring every attorney to file a certificate attesting either that no portion of a filing was drafted by generative AI or that a human checked any AI-drafted language for accuracy against print reporters or traditional legal databases (text as posted May 30, 2023). His current page says he no longer requires that certification, because the court adopted Local Civil Rule 7.2(f), which Special Orders 2-97 and 2-98 adopted, effective September 3, 2024 (judge page). Rule 7.2(f) says a brief prepared using generative AI "must disclose this fact on the first page under the heading 'Use of Generative Artificial Intelligence,'" and that a party whose brief has no such disclosure certifies that no part of it was prepared using generative AI (local civil rules).
• Fifth Circuit. In June 2024 the court decided not to adopt a special rule on AI in briefs, and reminded counsel that filings must be carefully checked for truthfulness and accuracy. Its notice says: "'I used AI' will not be an excuse for an otherwise sanctionable offense" (notice).
• California (from January 1, 2027). SB 574 (Chapter 858, Statutes of 2026), signed September 30, 2026, adds Business and Professions Code § 6068.1, which requires a lawyer who uses generative AI to disclose that use to the court "for all documents submitted to the court," and amends Code of Civil Procedure § 128.7 so that no filing contains a citation the responsible attorney has not personally verified, including one provided by generative AI. It has no urgency clause, so it takes effect January 1, 2027 (bill text).

Individual judges elsewhere have issued their own standing orders, with different requirements. Read the local rules and your judge's page for every court where you file. "The tool made it up" has not worked as a defense in the sanction orders above.

Jurisdiction-specific law

A general chatbot does not know your jurisdiction unless you tell it, and even then it can blend training data across states and years. Ask "what is the non-compete rule" and you may get a confident answer that is true in one state, wrong in another, and out of date everywhere. Do not rely on the answer without reading the statute or opinion for your jurisdiction.

Currency

A model's built-in knowledge stops at a training cutoff, and statutes and case law move. When ChatGPT search is on, it can pull in current web pages and show links to them (OpenAI). A link is something to open and read. It is not proof that the sentence next to it is right.

No source you can check by default

When ChatGPT answers from memory, the cite is a prediction and there is no underlying document it pulled from. The only way to verify is to find the source yourself, which is the step most people skip under pressure. When it answers with search results, you still have to open the page, confirm it says what the answer claims, and confirm it is a primary source.

The task-by-task verdict

The line is not "ChatGPT good" or "ChatGPT bad." It is whether the task needs the model to supply law it might invent, and whether the input carries client confidences.

The duties: ABA Formal Opinion 512

If you are a US lawyer, this is not a matter of taste. ABA Formal Opinion 512, issued July 29, 2024, applies the Model Rules to generative AI. Three points matter most for everyday ChatGPT use. Opinion 512 is ethics guidance. The binding rules are your state's, so check your state bar too, since several have issued their own guidance. See are ABA formal opinions binding.

Competence (Rule 1.1). Lawyers "must have a reasonable understanding of the capabilities and limitations" of the specific tool they use. Relying on or submitting a tool's output "without an appropriate degree of independent verification or review" could violate the duty of competence. The amount of verification depends on the tool and the task, and the opinion says a tool used for a discrete task such as generating ideas "may require less independent verification or review." Whatever level of review you choose, the lawyer is "fully responsible for the work on behalf of the client."

Confidentiality (Rule 1.6). The opinion says that, because many self-learning tools are designed so that their output could lead directly or indirectly to disclosure of information relating to the representation, "a client's informed consent is required prior to inputting information relating to the representation into such a GAI tool." It adds that "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient." As a baseline, lawyers should read and understand the terms of use and privacy policy of any tool they use. Where a lawyer is not inputting information relating to the representation, for example when using the tool for idea generation, informed consent is not required.

Candor and misleading filings (Rules 3.1, 3.3, 8.4(c)). The opinion says these rules may be implicated by certain uses. You stay responsible for everything you file.

We unpack the full opinion in plain English in our ABA Formal Opinion 512 guide.

Consumer vs business plans: the data distinction that decides what you can paste

This is the part most "ChatGPT for lawyers" pieces gloss over, and it governs your confidentiality duty. OpenAI's consumer plans and its business plans are different products with different data handling.

OpenAI's help page says that with ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu and its API Platform, "by default, we don't use provided inputs and outputs to train our models." For ChatGPT Plus, Pro and Free on a personal workspace, it says data sharing is enabled by default, and you can opt out under Your Profile, Settings, Data Controls, "Improve the model for everyone." OpenAI's enterprise privacy page (updated January 8, 2026) says it can execute a Data Processing Addendum for ChatGPT Business, ChatGPT Enterprise and the API. Always confirm the current terms yourself, because vendor policies change.

The two screenshots below show the consumer Data controls menu and the opt-out switch. OpenAI changes its menus, so follow the path on its help page.

OpenAI renamed the Team plan to ChatGPT Business as of August 29, 2025, and says it was a name change only (OpenAI).

A few practical notes. Pro is grouped with Plus and Free on OpenAI's help page, so do not assume the higher price buys business data terms. Turning off training in consumer settings is a setting, not a contractual commitment. On Business, OpenAI's privacy page says workspace admins can view, access, export and delete end-user conversations, so someone at your firm can read what lawyers type. And a court can override a retention policy: from May to September 2025, a federal court order required OpenAI to preserve output logs it would otherwise have deleted, including on consumer plans. We cover that order, and the 2026 ruling in United States v. Heppner that a defendant's chats with a public AI tool were not privileged, in is ChatGPT confidential for legal work and does ChatGPT have attorney-client privilege.

For the full picture of where your inputs travel and what to ask a vendor, see where your legal AI data actually goes.

When a research tool that links to sources helps

A general chatbot and a research tool built on source documents can look identical from the outside. Both are a text box. The difference is what happens before the answer appears.

A tool that retrieves real source documents, such as statutes or opinions, and answers with citations that link back changes the economics of your verification duty. With a chatbot answering from memory, checking a cite means rebuilding the research from scratch. With a source-linked tool, it is a click to the source text. That does not make verification unnecessary. You still read the source and confirm it says what the answer claims. It makes the check cheap. For how to run that check, see how to verify AI legal citations before filing.

So a reasonable division of labor:

• General chatbot: first drafts where you supply the law, rewrites, summaries of your own text, brainstorming.
• A source-linked research tool, or the primary source itself: anything that asserts what the law is and needs a citation a court can check, or confirming statutory text.

A simple workflow that keeps you safe

You do not have to choose one tool. Most in-house lawyers run both, with a clear handoff. The decision at each step comes down to two questions: is client-identifying data in the prompt, and does the output assert what the law is?

1. Draft and brainstorm in ChatGPT, supplying the facts and the law yourself, on a plan whose data terms fit the sensitivity of the matter.
2. Strip or anonymize anything client-identifying before pasting if you are on a consumer plan.
3. Pull and verify every legal proposition against the primary source, never from the chatbot's memory alone.
4. Treat the chatbot's output as scaffolding, not authority. You own every cited claim in the final document, and your court may require you to disclose AI use.

FAQ

Is it safe for lawyers to use ChatGPT? Yes, for the right tasks. It is safe for first drafts, summaries of text you provide, rewriting, and brainstorming. It is not safe as a source of legal authority, because it can fabricate citations. ABA Formal Opinion 512 expects you to understand the tool's limits and to verify its output to the degree the task requires.

Can ChatGPT do legal research? It can sound like it does, which is the trap. Answering from memory, it can produce confident citations with no real source behind them, including cases that do not exist. With search turned on it shows links, which you still have to open and read. For research you can cite, go to the primary source and verify the result.

Does ChatGPT use my data to train its models? On Free, Plus and Pro on a personal workspace, OpenAI says data sharing is enabled by default and you can opt out. For ChatGPT Business, Enterprise and Edu, OpenAI says it does not use inputs and outputs to train its models by default (OpenAI). Confirm the current terms, since they change.

Is ChatGPT Team still a plan? OpenAI renamed ChatGPT Team to ChatGPT Business as of August 29, 2025, and says it was a name change only (OpenAI).

What is ABA Formal Opinion 512? It is the ABA's July 29, 2024 ethics opinion on generative AI. It applies the Model Rules on competence, confidentiality, communication, candor and related duties to AI use. It asks lawyers to understand a tool's capabilities and limits, to verify output to an appropriate degree, to read the tool's terms, and to get a client's informed consent before inputting information relating to the representation into a self-learning tool.

Why does ChatGPT make up cases? Because it predicts plausible text rather than retrieving real documents. Asked for supporting authority, it can generate a citation-shaped string that looks correct. It has no reliable way to check whether that case exists, and it will not always tell you it is guessing.

ChatGPT or a legal research tool for in-house work? Often both. ChatGPT for drafting where you own the facts and the law, and the primary source, or a tool that links each cite to its source, for anything that has to cite the law accurately. The split is whether the task needs a checkable source or just good prose.

Is the ChatGPT Pro plan safe for client data? Pro is grouped with Plus and Free on OpenAI's help page, so do not assume the higher price buys business data terms. For client-confidential work, look at a business plan (Business or Enterprise) with a written data agreement, and read the terms before pasting anything sensitive.

Do I have to tell the court I used ChatGPT? It depends on the court. The Northern District of Texas requires a brief prepared using generative AI to disclose that on its first page (Local Civil Rule 7.2(f)). Other judges have issued their own orders, and some courts have none. From January 1, 2027, California law (SB 574) requires a lawyer who uses generative AI to disclose that use to the court on all documents submitted. Check the local rules and your judge's page before you file.

What is the best ChatGPT prompt for lawyers? There is no single prompt. The best ones share a structure: assign a role, paste the source material, and constrain the output format. Prompts that supply the facts and the governing text are safer. Prompts that ask ChatGPT to recall the law from memory are where fabricated citations come from.

Can ChatGPT replace lawyers? No. Opinion 512 says generative AI tools "cannot replace the judgment and experience necessary for lawyers to competently advise clients." It drafts language quickly, which frees a lawyer's time, but the lawyer still owns every judgment call and every cited claim.

Is ChatGPT confidential for legal work? Not on consumer plans by default, where OpenAI says data sharing for training is on unless you opt out. Business and Enterprise plans carry a no-training default and a DPA option, but Opinion 512 still asks you to read the terms and get informed consent where client information goes in. Chats are also not automatically privileged: in United States v. Heppner (S.D.N.Y. Feb. 17, 2026), the court held a defendant's chats with a public AI platform were not protected. See is ChatGPT confidential for legal work.

For the US statutes, regulations and court rules these laws sit in, see the US primary law API.

Arshita Anand

Co-Founder & CEO · Attorney

Arshita leads product and strategy at Vaquill AI, the US primary law data API. As an attorney, she decides which law we carry, how each section is sourced and cited, and what legal AI and compliance teams need from the data.