Is ChatGPT Safe for Lawyers? Confidentiality and Privilege

Is ChatGPT Safe for Lawyers? Confidentiality and Privilege

ChatGPT is safe for lawyers to use when no confidential client information goes into a consumer account and a lawyer checks every output before it reaches a client or a court. No ethics rule bans it: ABA Formal Opinion 512 says lawyers may use generative AI as long as they meet the duties they already have.

The catch is the setup. A free or Plus account can keep and learn from what you type, a 2026 federal ruling refused privilege for a defendant's own AI chats, and courts have fined lawyers for filing cases ChatGPT made up.

Below: what you can paste and what to keep out, what the ABA and state bars say, how privilege holds up, the settings to check this week, and safer setups for client work. This is general information, not legal advice, so check your own state bar's guidance before you change how your firm works.

Is ChatGPT safe for lawyers? The short answer

ChatGPT is safe for lawyers on public and non-client work, and not safe for confidential client information on a consumer plan. The tool is the same in both cases; what changes the risk is how your account is set up. Four things decide it:

  • Account type: Free, Plus and Pro are consumer plans. OpenAI says it does not train on data from its business plans (Business, Enterprise, Edu and the API) by default.
  • Training setting: consumer chats can be used to train models unless you turn off "Improve the model for everyone" under Settings, Data Controls.
  • Retention: OpenAI's stated policy deletes Temporary Chats and deleted chats within 30 days, but a court order can override a policy. In 2025, a preservation order in The New York Times' copyright suit made OpenAI keep consumer chats, deleted ones included, for several months.
  • Where it runs: a cloud chatbot processes your prompt on someone else's servers. A model running on hardware the firm owns keeps the prompt inside the office.

A quick test: if you would not email the text to an outside vendor without a signed agreement, do not paste it into a consumer chatbot.

What you can and cannot put into ChatGPT

Public information and generic drafting are fine on any plan; anything that identifies a client, a matter or your strategy belongs only in a tool with no-training terms you have vetted, or one the firm runs itself.

These tasks are fine on any plan, including a consumer account with training turned off:

  • Explain a statute or rule in plain language (public law).
  • Draft a blog post or newsletter (firm information only).
  • Tidy a clause template once names, facts and deal terms are removed.
  • Find case law for a brief, as a lead only: citations can be invented, so verify every cite on any plan.

Tasks that touch confidential client information need a different setup:

Task with client dataConsumer ChatGPTBusiness plan, no trainingModel on firm hardware
Summarize a client contract or depositionNoAfter vetting terms; consent may be neededYes, under firm access rules
Brainstorm case strategy (privileged)NoRisky; keep it lawyer-directed and loggedYes, under firm access rules
Draft a client letter with case factsNoAfter vetting terms; review every lineYes; review every line

Removing names helps less than it looks. A rare fact pattern, a deal size or a hearing date can point to one client even without a name, so strip those too or keep the task out of consumer tools.

What the ethics rules say about lawyers using ChatGPT

No ethics rule names ChatGPT, but the existing rules on competence, confidentiality, supervision, candor and fees decide how you may use it. The ABA and several state bars have now spelled out how those rules apply.

Colleagues at a shared table working on laptops and reviewing papers

The ABA Model Rules that apply

These Model Rules carry most of the weight. Your state's version may differ in wording, so treat these as the common baseline:

  • Rule 1.1, Comment 8 (competence): lawyers keep up with "the benefits and risks associated with relevant technology". Most states have adopted some version of this comment.
  • Rule 1.6(c) (confidentiality): lawyers make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or access to, client information.
  • Rules 5.1 and 5.3 (supervision): partners and supervising lawyers answer for how the firm and its staff use tools, AI included.
  • Rules 3.1 and 3.3 (meritorious claims and candor to the court): what you file must be accurate, which is why checking AI output is a duty, not a courtesy.
  • Rule 1.5 (fees): what you bill must be reasonable: bill the time you actually spent, not hours the AI saved.
  • Rule 1.4 (communication): lawyers consult the client about the means used to reach the client's goals, which can include telling them how AI is used on their matter.

ABA Formal Opinion 512 in six points

Formal Opinion 512, issued July 29, 2024, is the ABA's main guidance on generative AI in law practice. It is advisory, not binding, but many state bars cite it. Its core points:

  1. Understand the tool. You do not need to be an AI expert, but you need a "reasonable understanding of the capabilities and limitations" of the tool you use.
  2. Weigh the disclosure risk first. Before entering client information, assess whether the tool could expose it.
  3. Get informed consent for self-learning tools. Client information should not go into a tool that learns from inputs without the client's informed consent, and a boilerplate line in an engagement letter is not enough.
  4. Verify the output. Do not rely on AI output without an appropriate degree of independent review.
  5. Supervise. Managing lawyers set policies, train staff and check that the policies are followed.
  6. Bill fairly. You can bill the time you actually spend, including review, but not the time it takes to learn a tool you will use for clients regularly.

State bar guidance: California and Florida

California and Florida both allow generative AI and both put confidentiality first, but they draw the consent line in slightly different places.

California: the State Bar Board of Trustees approved updated Practical Guidance on May 14, 2026, replacing the November 2023 version and adding agentic AI.

It says a lawyer must not input confidential client information into a tool that may present material risks to confidentiality or security without the client's informed consent.

Reasonable efforts, it adds, "require more than reliance on generalized marketing assurances": read the terms of use and privacy policy, and bring in IT or security help where needed. AI agents with access to email or document systems should get only the access they need.

Florida: Ethics Opinion 24-1 (January 19, 2024) recommends informed consent before using a third-party AI tool that would involve disclosing confidential information, and tells lawyers to research the tool's policies on data retention, data sharing and self-learning.

It also notes that confidentiality concerns may be reduced by an in-house AI where the data is not hosted by a third party, and that no consent is required under its Rule 4-1.6 if no confidential information goes to a third party. Florida also says client-facing chatbots must say they are AI, not a lawyer.

Other states have issued their own opinions, and they do not all agree on when consent is required. Look up your bar's guidance before you settle on a firm policy.

Does ChatGPT waive attorney-client privilege?

Using ChatGPT does not waive privilege by itself, but confidential material shared with a public AI platform can lose protection, and in an early federal ruling, United States v. Heppner, the person who shared it lost.

Three colleagues talking beside a standing desk in a bright office

In United States v. Heppner (S.D.N.Y.), Judge Jed Rakoff ruled from the bench on February 10, 2026 and explained his reasons in a written opinion on February 17.

About 31 documents that a criminal defendant created with Claude, Anthropic's chatbot, were held not protected by attorney-client privilege or the work product doctrine. He made them on his own, not at his lawyers' direction, and later shared them with counsel.

The court's reasons, as reported by the Harvard Law Review blog and BakerHostetler:

  • Claude is not a lawyer, so there was no attorney-client communication to protect.
  • No reasonable expectation of confidentiality: the platform's privacy policy allowed it to collect inputs and outputs, use them for training and disclose them to third parties, including in litigation.
  • Sending the outputs to counsel later did not make them privileged.
  • No work product: the documents were not prepared by or at the behest of counsel.

What the ruling did not decide matters just as much. It concerned a client acting alone, not a lawyer.

The court suggested the result might differ if counsel had directed the AI use, since the tool could then act as the lawyer's agent. The pleadings did not show the defendant's privacy settings, and the ruling did not address enterprise tools with contract-level confidentiality. Commentators also disagree with parts of its reasoning.

The law is not settled. On the same day, a magistrate judge in the Eastern District of Michigan went the other way on work product in Warner v. Gilbarco, Inc. (No. 2:24-cv-12333), a civil case. He held that a self-represented plaintiff's ChatGPT materials were protected work product, calling AI programs "tools, not persons", so using them did not hand anything to the other side.

The law-firm alerts that followed give similar advice: check a tool's confidentiality terms before use, prefer closed enterprise systems, have counsel direct AI work on legal matters, and write an AI-use policy. One step most firms can take today is to tell clients, at intake, that running case facts through a public chatbot on their own may not be privileged.

Fake citations: the accuracy risk behind the sanctions

The best-known AI sanctions case is Mata v. Avianca, where a federal judge in New York fined two lawyers and their firm $5,000 in June 2023 for filing judicial opinions that ChatGPT had invented.

Two colleagues working at standing desks in a white office

In his June 22, 2023 order, Judge P. Kevin Castel wrote that there is "nothing inherently improper about using a reliable artificial intelligence tool for assistance", but that the rules "impose a gatekeeping role on attorneys to ensure the accuracy of their filings."

The lawyers were sanctioned under Rule 11 because they filed fake opinions with fake quotes and kept standing by them after the court questioned whether they existed.

Besides the penalty, they had to send the order to their client and to each real judge falsely named as an author.

The lesson is not "never use AI for research". It is that the signature on the filing is yours. Before anything AI-assisted goes to a court:

  • Pull every cited case in Westlaw, Lexis or the court's own records and read the page you cite.
  • Check every quote word for word against the source.
  • Read the judge's standing orders: some courts require you to disclose or certify AI use in filings.

Settings and contracts to check before using ChatGPT for client work

Six checks cover most of the confidentiality risk, and the first two take a few minutes:

  1. Training off: on any consumer account, turn off "Improve the model for everyone", and use Temporary Chat for anything you do not want kept in history.
  2. No shared links: a shared chat link shows the conversation to anyone who has the URL. Share drafts through your document system instead.
  3. The right plan: move client work to a business plan whose terms exclude your data from training, and read the retention terms, not only the marketing page.
  4. A data agreement: for client data, get the vendor's data processing terms in writing and keep a copy with your policy.
  5. Logging and access: know who in the firm can use the tool, with what data, and where chats are stored.
  6. Client consent: where your bar or the client's own policy requires it, get specific informed consent before client information goes into a third-party tool. Some corporate clients set AI rules in their outside counsel guidelines, so check those too.

The same checks apply to any cloud AI service, not only ChatGPT. Our guide to AI privacy and security covers the general data risks in more detail.

Open-plan office with staff at their desks and an Autonomous Computer on a filing cabinet

Safer ways for lawyers to use AI with client data

There are three safer setups for client work: a business plan with no-training terms, a legal-specific AI tool, or an open model running on hardware the firm owns. Each one reduces a different risk, and none removes your duty to supervise and verify.

Business and enterprise plans with no-training terms

A business plan is the fastest step up from a consumer account. OpenAI says it does not train on data from ChatGPT Business, Enterprise, Edu or the API by default, and admins get controls over retention and access.

Your prompts still go to the vendor's servers, so the contract terms are what protect you. Read them, and keep in mind that a court order can still reach data a vendor holds.

Legal-specific AI tools

Legal AI products wrap a model in features lawyers need, such as citation links to real authority, document management connections and matter-level permissions. They are still cloud services in most cases, so the same vetting applies: training terms, retention, where data is processed and who at the vendor can see it.

What an AI legal assistant does, and how to compare one against another, is a separate question from whether it is safe.

An open model on hardware the firm owns

Running an open-weight model on a machine in your office keeps prompts and documents inside the building, which is the setup Florida's opinion describes when it says in-house AI can reduce confidentiality concerns.

You trade convenience for control: someone has to run the machine, and open models answer differently from the newest cloud models, so test them on your own non-confidential samples first. If you are new to the idea, start with how to run AI locally.

Autonomous Computer 2 next to a monitor on an office desk, running AI models locally

The Autonomous Computer line is built for this. Autonomous describes it as "private by physics": prompts never leave the building, nobody trains on your work, and the model weights sit on your own disk, so a model is not deprecated or rate-limited by a vendor.

It ships with Ubuntu, NVIDIA drivers, vLLM and SGLang preinstalled. Computer 2 is the personal AI computer that sits on a desk, which suits a solo attorney or a small practice.

Computer 4 is the business AI computer, made to run on-prem for a team that shares one machine.

Both machines come with a 3-year warranty and ship in 4-6 weeks.

One honest limit: on-prem moves where the data lives, but it does not make a firm compliant by itself. Autonomous makes no compliance or certification claims for these machines. Access control, encryption, backups, audit logs and staff training are still the firm's job, and the duty to check every output stays with the lawyer.

Keeping chats off third-party servers is covered in this explainer on private AI chat.

Using ChatGPT for lawyers: low-risk tasks

ChatGPT is low-risk for lawyers on tasks that use only public or firm information and never touch a client's facts. These are good places to start using AI in a law firm, even on a consumer plan with training turned off:

  • Research starting points: ask for the main issues in an area of law or a list of terms to search, then do the real research in a citator.
  • Plain-language explanations: turn a public statute or a court process into a short explainer for clients.
  • Marketing drafts: blog posts, bio updates and newsletter outlines, checked against your bar's advertising rules.
  • Writing help: tighten your own non-confidential prose, outline a CLE talk or draft an internal checklist.
  • Templates: a generic clause or letter skeleton with no client details, which you then adapt inside your own systems.

Treat every answer as a first draft from a fast but unreliable assistant. It saves typing, not judgment.

FAQs

Is it illegal for a lawyer to use ChatGPT?

No, it is not illegal for a lawyer to use ChatGPT. ABA Formal Opinion 512 and state bars such as California and Florida allow generative AI, as long as the lawyer protects client confidentiality, checks the output, supervises staff use and bills fairly. Breaking those duties, not using the tool, is what leads to discipline or sanctions.

Is it safe to put legal documents in ChatGPT?

It is not safe to put confidential legal documents into a consumer ChatGPT account. Public documents such as statutes or published opinions are fine. Client contracts, pleadings drafts or discovery belong only in a tool whose terms exclude training and that you have vetted, or in a model the firm runs on its own hardware.

Can ChatGPT chats be used in court?

Yes, ChatGPT chats can be used in court, because they are records like emails or notes. They can be requested in discovery or seized, as the AI chats in the Heppner case were. Deleting a chat does not guarantee it is gone: in 2025 a court order made OpenAI keep deleted consumer chats for months.

Does using AI break attorney-client privilege?

Using AI does not automatically break attorney-client privilege, but confidential material shared with a public AI tool can lose it. In United States v. Heppner, a client's own chats with Claude were held not privileged. Lawyer-directed use of a tool with real confidentiality terms stands on firmer ground, though courts have not settled it.

Are there any lawyers caught using AI in court?

Yes, lawyers have been sanctioned for filing AI-generated fake citations. The best-known case is Mata v. Avianca, where a New York federal judge fined two lawyers and their firm $5,000 in 2023. Courts have handled many similar cases since, so verify every citation yourself before filing.

Is Claude or ChatGPT better for lawyers?

Neither Claude nor ChatGPT is safer for lawyers by brand; the plan and its data terms matter more than the model. Both makers offer business tiers that exclude your data from training by default. For drafting quality, test both on the same non-confidential tasks and keep the one whose output needs less fixing.

Do I need client consent to use ChatGPT?

ABA Opinion 512 requires informed consent before client information goes into a tool that learns from inputs, and Florida's Opinion 24-1 recommends consent before any third-party AI tool sees confidential information. Consent must be specific, not boilerplate. Check your own bar's rule; tasks with no client information need none.

Can lawyers bill clients for time spent with ChatGPT?

Lawyers can bill the time they actually spend using ChatGPT on a matter, including reviewing its output, under ABA Opinion 512. In most cases they cannot bill for time spent learning a tool they will use regularly, and Florida's opinion says AI does not permit improper billing such as double-billing.

Conclusion

ChatGPT is safe for lawyers on public, non-client work today, with training turned off and every citation checked.

Client information needs more: a vetted business plan, a legal tool with clear terms, or a model on hardware the firm controls, plus the client's informed consent where your bar requires it. Start with the settings checklist this week, write a one-page AI policy for the firm, and read your own bar's guidance before client data goes anywhere new.

References

‌