ChatGPT for Lawyers: What Rule 11 Still Requires in 2026

TL;DR: ChatGPT for lawyers works well for a narrow set of jobs: drafting, summarizing, and brainstorming. It is not safe as a legal research tool.

Rule 11 of the Federal Rules of Civil Procedure puts the duty on the signer, not on the AI. In Mata v. Avianca, Judge Castel sanctioned two lawyers and a firm. Their brief carried fake cases from ChatGPT.

The working rule is simple, and most firms can set it in a week. Treat the model like a new law clerk. Check the work line by line before it goes out.

The topic is a working one now, not a futurist one. Judges have ruled on it, state courts have written rules, and firms have fired people over it.

This article covers what ChatGPT is and what courts expect of lawyers who use ChatGPT. It also covers the practical rules from the first wave of AI cases.

The phrase shows up in search as both "chat gpt for lawyers" and the one-word form. This article covers both.

What does ChatGPT do for lawyers today?

ChatGPT is a general chat tool from OpenAI. It writes text from a prompt. ChatGPT is one generative AI tool among many.

Lawyers use ChatGPT today for drafting, not for case research. Firms use it to produce letters, to summarize long records, and to brainstorm issues.

The model is one of many generative AI tools in use at law firms. AI tools like ChatGPT help lawyers with first-version drafts.

The vendor did not build it for case research. A person who uses it on a client matter is still doing the practice of law.

The ethics rules do not change, and the civil rules do not change either. What changes is the speed of the first draft, and the signer must now catch a new type of error.

The duty still runs to the signer, even when a program wrote the first version. The use case is drafting with a check, not research without the check.

ChatGPT does three jobs well in legal writing. It drafts prose fast, it summarizes long files, and it moves text between formats. A clean example is turning case notes into a client letter.

These three jobs cover a large share of daily work at many firms. ChatGPT in legal use is a drafting partner, not a research service. The jobs do not cover case research, and that is where the risk lives.

A core AI point sits under all of this. The model does not know if what it wrote is true. The signer still owns every claim in the paper.

What happened in Mata v. Avianca?

Mata v. Avianca is the case that put AI in a law firm in the national press. In 2022, Roberto Mata sued the airline Avianca after a metal cart hit his knee.

His lawyers filed a brief that cited several cases. The cases did not exist, and the other side could not find any of them.

A member of the Levidow firm used the model to find supporting cases. He pasted the AI outputs into the brief. The other side could not match the references to any of the trusted legal databases.

The court asked for copies. The lawyer asked ChatGPT again. The AI then made up excerpts that also looked like real court opinions.

On June 22, 2023, Judge Kevin Castel of the Southern District of New York ruled. He wrote a 43-page opinion on sanctions.

He wrote that existing rules place a gatekeeping role on lawyers to ensure accuracy. He found that two lawyers and a law firm acted in bad faith.

The court imposed sanctions under Rule 11. It fined the lawyers and the firm $5,000 jointly. The firm shared the liability.

The court also made them send the opinion to the real judges. The AI had attached those judges' names to fake cases. The case became the shorthand for what happens when a signer skips the check.

The lesson is not about using the model. The lesson is about skipping the check at the signer stage.

For a longer read, see the Mata sanctions regulatory regime piece. It covers how Mata led to the rules that followed.

Rule 11: the gatekeeping role every signer still owns

Rule 11 of the Federal Rules of Civil Procedure is the rule the Mata court used. It sets out what a lawyer promises on every court paper. Those promises have not changed since the model came along.

What has changed is how courts test those promises in day-to-day cases. The rule rests on the signer, and it never rested on the tool.

Rule 11(b) sets out four certifications. By signing or filing a paper, the signer certifies to the best of what they know after a reasonable inquiry.

The paper is not filed for an improper purpose. The legal arguments rest on existing law or on a nonfrivolous argument for a change.

The factual claims have evidence or will likely have evidence after a reasonable look. The denials of fact rest on the evidence or on a lack of information.

The duty runs to the signer, and the firm shares it. The scope has not shrunk since the rule was last amended.

Rule 11(c) sets up the sanction side. The court may sanction a lawyer who breaks the promise, after notice and a chance to respond.

Sanctions can include fines, striking the paper, and an order to pay the other side's fees. A firm is on the hook for its member's work under subsection (c)(1).

The rule does not name any tool, and it does not need to. The question is whether the lawyer did a reasonable inquiry under the facts.

A person who files cases from ChatGPT without a database check did not do that inquiry. The Mata opinion made that point in plain terms.

AI for lawyers: what the tool can do safely in a law firm

The main job for the model in a law firm is drafting with a review step. The signer still writes the final paper. AI for law firms is a tool bucket, not a lawyer replacement.

This is the first lane where the model earns its keep. ChatGPT outputs that look confident still need a check.

ChatGPT can produce a demand letter from a set of bullet points. The lawyer reads it, cuts what does not fit, and signs the final version.

The model can turn a 40-page deposition summary into a one-page memo. The reader still reads the deposition itself for anything that will go in a court paper.

The tool can also prepare a contract clause from an outline. A contract review step still runs on the clause before it goes to the other side.

Legal AI tools handle several other short jobs well. The model can write a plain-English version of a complex paragraph for a client.

It can produce an outline for a brief from an issue list. Legal brief writing is one of the places where the model earns its keep. It can also summarize a long email thread into a short recap for in-house legal teams.

The use of AI on these tasks does not touch case research. The signer only writes the paper and runs the check on the output.

ChatGPT can convert notes into a client letter or brainstorm arguments. It can help a legal assistant draft an intake memo. Many firms already use it for intake notes, calendar entries, and short internal emails.

These are low-stakes legal tasks where a person reads the output before anything leaves the firm. That is the general shape of safe practice.

For a case-by-case look at the daily work this covers, see how a working lawyer uses Claude. The pattern is the same with any chat tool.

ChatGPT and confidentiality: what you put in matters

The second place AI use at the firm can go wrong is privacy. The civil rules handle what leaves the firm as a court paper. The Rules of Professional Conduct handle what leaves the firm as information.

The model sends what the lawyer types to the vendor's servers for processing. That matters for client secrets and for privilege.

Model Rule 1.6 and its state kin require that lawyers keep client information confidential. The duty covers the tools used in the work. Prudent practitioners treat the model with those limits in mind.

A person who pastes a full client name and case facts into a free chat tool has done something risky. The information has gone to a third-party vendor.

OpenAI publishes a data use page that explains the account types. Consumer accounts may use chats to train future models unless the user turns training off in settings.

Business and enterprise accounts do not use inputs for training. The API does not use inputs for training either, by default.

AI features in paid tiers change the model, not the signer's duty. The product rules change often, so the policy page is the source of truth.

A person who relies on the model should read the current data use page. Check it again every few months, and write the rule into a short firm policy.

The safe habit is to strip identifying details before any prompt. Names come out, case numbers come out, and addresses come out.

The lawyer can still ask the model for help with the shape of the problem. The person does not have to feed the tool the client file.

A redaction habit is cheap, and a privilege waiver is not. The daily practice version lives in the solo AI stack privacy piece.

For the court orders on AI and privilege, see the Heppner ruling. The companion ruling lives at the Assini v. Hayward piece.

Can the model do legal research?

The honest answer is partly yes and mostly no. ChatGPT for legal research produces fake cases that look real. A person can use the model to shape a research question and to outline the issues.

They can also brainstorm the areas of law that may apply. Lawyers need a database check on every ChatGPT citation before it goes in a legal document.

They cannot use the tool to find cases to cite without an independent database check. The Mata ruling is the reason this rule is now a hard one in federal court.

The model does not have a database of real cases behind it. It has patterns of language, and case names and reporter numbers follow strong patterns the system learned.

ChatGPT can produce text that looks like a real cite when the case does not exist. The output reads correctly, but it is still wrong on the facts.

The judge or the other side will catch it, and the signer is on the hook under Rule 11(c)(1).

At the research stage, treat the model's output as leads, not as sources. Each lead gets a check in Westlaw, Lexis, Fastcase, Google Scholar, or the court's own website. Dedicated legal research tools exist for a reason.

If the reader cannot find a lead, the reader cuts it from the draft. If the reader finds it, the case still gets a full read. The model also mischaracterizes real cases.

The simple rule works best. If the model gave you the cite, you did not research it yet.

The research step is pulling the case, reading the case, and confirming the holding. All three steps take a trained reader, not a chat tool.

For a longer walk, see why I stopped asking the model for cites.

The hallucination problem: why fake cites keep happening

Hallucination is the AI term for an output that is confident and false. In a court paper, the most common form is a fake cite that looks real on the page.

The judge or the other side will catch it, and the signer is on the hook. The firm shares the risk under Rule 11.

A fake cite is not a bug the next version will fix. It is a side effect of how large models work in day-to-day use.

The model picks the most likely next word based on what came before. For a legal cite, the next words are a plausible case name and a plausible reporter number.

The model does not stop to ask if the case exists. It fills in a pattern that looks right.

The failure shows up most in two places during AI-assisted drafting. The first is a narrow issue the model has not seen often. The second is a case summary by name.

For the narrow issue, the model tends to invent cases that fit the question. For the named case, the model tends to shape the holding to fit the question.

Both outputs are easy to miss on a quick read. The signer is responsible for every reference in the paper.

The courts have already written rules around this problem. Connecticut asks anyone who used generative AI to create or edit a filing to verify the AI output.

Federal courts rely on Rule 11, and the Second Circuit applied it in Park v. Kim (2024). A lawyer who did not check AI-produced cites violated Rule 11 under that decision.

For the state rules, see state court AI rules in Florida, New York, and Connecticut. For the federal cases on intent, see the Jarbou DOJ post on intent.

How do you check what the model wrote before you file?

The check is the step that turns a risky draft into a safe paper. The check is not optional, and courts do not look kindly on a shortcut.

Rule 11 asks whether the inquiry was reasonable under the facts. A sound inquiry for AI-assisted writing has a shape. Most lawyers can learn the shape in an afternoon.

Pull every cite yourself from a database you trust. If the model produced a brief that lists five cases, pull all five before any paper goes out.

Confirm the case exists, confirm the court, and confirm the year. Then confirm the holding the model gave to the case.

The signer cuts any reference that fails this test. The signer either finds a real case for the point or drops the point.

Read the case itself, not just the headnote. The model sometimes cites a real case for a holding the case does not support.

A quick read of the opinion catches that pattern. A paralegal or junior associate can do the read.

The tool cannot do the read, and the duty sits with the signer. The firm shares the duty under Rule 11.

Keep a record of each AI-assisted draft. Save the chat log that produced the first version, and note which paragraphs came from the model.

Note which paragraphs the lawyer added. Note what the person checked and when the check happened. The entry also serves as a training trail for the next legal professionals on the matter.

A one-paragraph entry in the file is enough. The record protects the signer if the paper is later questioned. The question of legal accuracy still rests on the signer, not on the tool.

For a step-by-step frame, see verification frameworks for AI-generated writing. It covers the same check with more detail.

A simple workflow for lawyers who already use ChatGPT

A workflow is a habit, not a product. Lawyers who use ChatGPT today can set a simple pattern. The pattern keeps the work inside Rule 11 and the ethics rules.

Generative AI use at the firm starts with a short written policy. AI adoption in the legal field runs on a policy and a check.

The first step is prompt discipline on every chat. Describe the task without client-identifying facts. Ask for a shape of the paper, not a final version.

Ask for a first version that cites nothing, and let the lawyer add cites from a legal database. That one habit cuts the fake-cite risk at the source.

The second step is a line-by-line read before anything leaves the firm. The reader scans the paper like a first-year associate's work.

Any claim about the law gets a check against a legal database. Any factual claim about the client's matter gets a check against the client file.

The reader pulls every reference from a database. Anything that reads too confident gets a rewrite from the signer.

The third step is a short note in the client file. The note names the AI tool and the general purpose of the chat.

It also names who did the review and when. A one-paragraph entry is enough. AI assistance on a court paper is a trail the firm wants on record.

The court may later ask what the firm did on an AI-assisted paper. The firm now has an answer that fits what Rule 11 expects of a signer.

The fourth step is a quarterly habit audit at the firm level. The practice management team walks through its last ten AI-assisted filings.

The team asks what went well and what almost missed. Patterns show up fast in a sample that size.

A clause that keeps slipping through goes on the standing checklist. A cite style that keeps getting invented gets a flag for closer review.

The audit also catches when the firm has drifted from its own written policy. AI ethics at the firm is a habit, not a one-time memo.

Common questions about ChatGPT in a law firm

Can a lawyer use the tool at all?

Yes, and no court has banned it in a filing. The rules put the duty on the person who signs the paper, not on the AI.

Is ChatGPT confidential?

Not by default. Consumer accounts may use inputs to improve the model, and the user can turn training off in settings. Business and enterprise accounts and the API do not use inputs for training by default.

Firms should still strip client identifiers before any prompt. A short written policy helps the whole legal field. OpenAI's data use page is the source of truth, and the page can change.

Can the model advise a client?

No. ChatGPT is a general AI product. Only a lawyer in a client relationship can advise a client.

The output is a draft or a summary. The signer still owns the words that leave the firm.

Which tasks does the tool handle well?

Draft prose, summaries of long documents, and outlines of briefs. Also format changes, brainstorming arguments, and first-draft client letters.

These are workflow tasks the lawyer can review quickly. The signer still produces the final version that goes out the door.

Which tasks does the tool handle badly?

Legal research that produces case cites without a database check. Also factual claims the reader cannot check against a source.

Any work that involves private client information in a consumer account is risky. These are the tasks that drove the sanctions in Mata and the cases that followed.

Do I have to tell the court I used an AI tool?

It depends on the court.

Florida's statewide rule focuses on accuracy and bars local courts from adding their own AI disclosure rules. Connecticut asks anyone who used AI to create or edit a court paper to verify the output.

What does Rule 11 require from a firm that uses the model?

A reasonable inquiry before signing. In practice, that means an independent check of every reference and every legal point.

The firm shares liability when the signer skips the check. The duty does not shift to the AI, and it never does.

Where should a firm start?

With a short written policy. The policy lists the legal tools lawyers may use, and it names which account types the firm permits.

It also names what information cannot go into any prompt. It describes how the firm checks AI-assisted drafts.

Rules in this area are moving fast across state and federal courts. Legal standards here come from state rules, federal rules, and bar guidance.

The best working habit is simple, and most firms can set it in a week. Lawyers and legal teams run the check together.

The model earns its keep when the person treats it as a drafting partner. The work gets a check. The AI fails when the person treats it as a database.

The first path is what quietly runs in firms every day. The second is what put the lawyers in Mata in the news.

Everything in this article is general information. Court rules and bar rules change often, so the signer still owns the final paper.