A lawyer's duty of confidentiality does not stop at the office door. It covers what you type into ChatGPT, Claude or Gemini. This article explains what that duty covers at a general level, what the American Bar Association said about AI tools in 2024, and a working method that keeps the client out of the prompt.

A leather-bound book, a legal pad and a navy folder on a wooden desk.

What the duty covers

Under ABA Model Rule 1.6, a lawyer keeps confidential all information relating to the representation of a client, regardless of its source. That is wider than "secrets". A client's name, the facts of the dispute, a draft settlement figure and the other side's identity all count, even when some of it is public elsewhere.

Rule 1.6(c) adds that a lawyer must make reasonable efforts to prevent unauthorized access to, or disclosure of, that information. Typing client material into a third-party AI service is a disclosure to that service. Whether it is a permitted one depends on the tool, the terms and what you typed.

The Model Rules are a template. Each state adopts its own version, and outside the United States the rules differ again. Check the rules of your own jurisdiction and your bar's ethics guidance before you rely on anything here.

What ABA Formal Opinion 512 says

In July 2024 the ABA issued Formal Opinion 512 on generative AI tools. In short:

  • Competence (Rule 1.1). You need a reasonable understanding of what the tool can and cannot do. Relying on its output without an appropriate degree of independent review can breach the duty of competence.
  • Confidentiality (Rule 1.6). Before you put information relating to a representation into a self-learning tool, you need the client's informed consent. The opinion explains the concern: what one client's matter teaches the tool may surface later in answers to other users.
  • No consent needed without client information. The opinion says consent is not required when you use a tool without inputting information relating to the representation, for example for idea generation.
  • Communication and fees. Clients may need to be told about AI use in some situations, and the way you bill for AI-assisted work has to stay reasonable.
  • Supervision. Partners and managers are responsible for setting AI policies for the lawyers and staff they supervise.

The opinion also says its conclusions rest on the risks and capabilities of these tools as of its publication. Those can change, and so can the guidance.

Your own bar may say more

Several state bars have published their own opinions on AI, and they do not all say the same thing. Some focus on consent, some on supervision, some on billing. Read the one issued by the bar that licenses you. If it is silent, Opinion 512 is a reasonable reference point, but it is not binding on you.

Privilege is a separate question from confidentiality. Whether sharing material with an AI service affects the privilege depends on your jurisdiction and the facts, and it is worth raising with your bar's ethics counsel before you rely on any assumption.

The practical way out: keep the client out of the prompt

The opinion's own example points to the simplest route. If what you type contains no information relating to the representation, the consent question does not arise. A clause can be reviewed without knowing who signed it. A timeline can be tightened without the client's name.

In practice that means replacing identifying details before anything leaves your computer. "Maria Lopez v. Northwind Ltd" becomes "Person_1 v. Company_1". The AI still sees who did what to whom. If you replace each name the same way every time, it can follow the argument across pages. The article on placeholders versus black bars explains why that matters.

Be honest about the limit. Names are only one part of "information relating to the representation". A rare fact pattern, an unusual amount or a small town plus a date can still point to your client. Replacing names lowers the risk and may change the analysis, but it does not make the question disappear. For matters where the facts themselves are the secret, either get informed consent or do not use a general-purpose AI tool.

A workflow you can defend

  1. Decide whether the task needs client facts at all. Drafting a template, explaining a doctrine and restructuring an argument often do not.
  2. Check the account. Know which plan you use and what its data settings are. The article on AI data settings lists what the vendors' help pages say.
  3. Send the passage, not the file. Take only the part you need.
  4. Replace every identifying detail. People, companies, addresses, account and case numbers, dates of birth, and the other side. See the checklist of what to remove.
  5. Read the safe version before you send it. One missed spelling of a name is enough.
  6. Check the answer. Verify every citation and every statement of law yourself. Then restore the real names locally.
  7. Keep a record. Note what you did and why, in case the client or your bar asks.

Frequently asked questions

Do I need client consent if I replace the names?

Opinion 512 ties the consent requirement to inputting information relating to the representation. If what you send no longer contains such information, the requirement does not apply on its face. Whether your version is clean enough is a judgment call, and your jurisdiction may take a stricter view.

Is a paid business plan enough?

A business plan that does not train on your data reduces one risk. It does not remove the fact that the material sits with a third party. Treat it as one layer, not the whole answer.

Can I tell the client I used AI?

In some situations you may have to, and in others it is good practice. An engagement letter that explains how you use AI tools is a sensible place to say it.

Who is responsible if the AI is wrong?

You are. The tool's output is a draft. Opinion 512 says independent review of that output is part of the duty of competence.

ClientSafe is a Windows app that replaces names, companies, addresses and ID numbers with consistent placeholders on your PC and puts the real names back into the AI's answer. It shows you what it found so you can correct it before anything is copied. Detection is not perfect, which is why the review step exists.

General information, not legal advice. Check the rules that apply to you.