Can lawyers ethically use AI?

Solo and small-firm lawyers ask this question with a specific worry behind it: bar discipline, or a sanctions order with their name on it. The short answer is that the profession’s regulators have already answered, in writing, and the answer is yes with conditions. This page collects what they said, turns it into five duties in plain firm terms, and closes with a setup a small firm can finish in one afternoon.

What the regulators say

No United States bar regulator that has spoken on the question bans lawyers from using AI. The pattern in every document this page checked is permission with conditions: use the tool, and stay inside the duties you already practice under.

The American Bar Association set the frame in July 2024 with Formal Opinion 512. It applies the existing Model Rules rather than writing new ones: competence with the tools the firm uses (Model Rule 1.1), protection of client confidences (Rule 1.6), and supervision of the work the way a firm supervises any assistant (Rules 5.1 and 5.3), with communication and fee duties alongside. A growing number of state bars have added their own guidance on top. Three documents give the shape of it.

Guidance moves. California replaced its 2023 practical guidance with a 2026 revision that covers autonomous AI tools, so check the current version of your own state’s document on the day you decide, and note the date. The verdict underneath is steady: no regulator wrote a new rulebook for AI, and every one of them applied the rules firms already practice under. That is why the rest of this page is organized by duty rather than by tool.

The five duties in plain firm terms

Across the opinions, the same five duties decide the ethics question. Each item below links to the step on this page that puts it into practice, because a duty the firm has answered in writing is a duty it can defend.

Write the answers down. An ethics posture that lives in one partner’s head does not survive a busy week, and written answers are what a bar inquiry asks to see.

Confidentiality does the heaviest work

The duty of confidentiality is where AI use most often goes wrong, and the regulators treat it that way. California’s guidance states it plainly: a lawyer must not put confidential client information into a generative AI tool that presents material confidentiality or security risks without the client’s informed consent. It also names the standard for vetting a tool, and marketing pages do not meet it. Reasonable efforts means reading the terms that apply to your plan: the terms of use, the privacy policy, the vendor documentation.

That reading is tool-specific work. For Claude, on commercial plans, Anthropic states customer data is not used to train its models by default, while on personal plans that choice sits in the account holder’s settings. The walkthrough of those terms, quote by quote, lives on a page of its own: Is Claude safe for lawyers? Whatever the tool, the defensible pattern is the same. Run the first tests on a fictional matter, fake firm, fake clients, real workflows, and let client data in only after the firm has written down what may go in. Control what goes in, and the confidentiality duty stays where the firm can manage it.

Verification is the duty courts see

Generative AI can produce a confident answer that is wrong, and it can cite a case that does not exist. That failure has already reached the courts: in a widely known 2023 case in a federal court in New York, lawyers were sanctioned for filing a brief that cited judicial opinions the tool had invented, and Texas Opinion 705 cites that case as a warning.

The regulators answer it the same way. Texas says a lawyer should always verify the accuracy of any response received from a generative AI tool. California says review every output for accuracy, citations to authority included, before it goes to a court, and correct any error it contains. The 2026 California revision draws one more line for tools that act on their own: no document reaches a court without a lawyer’s review and approval, and an AI system never files or speaks to a court on a lawyer’s behalf.

For a firm, the practice is short enough to memorize. Review-ready, never send-ready. A licensed attorney approves everything before it leaves the firm, and every citation gets checked against its source before anything is filed.

Disclosure and fees, the two quiet duties

Client communication comes first. California tells lawyers to consider disclosing AI use to the client, including how the technology is used and what its risks are, and Formal Opinion 512 points to informed consent before confidential client information goes into a self-learning tool. Many firms settle this once with a standing paragraph in the engagement letter, which is exactly the decision step 4 below records.

Fees are the quieter trap. The rule the opinions share: bill for the time you spent, the prompting and the reviewing included, and never for the time the tool saved. Texas Opinion 705 says a lawyer may not charge hourly fees for time saved by generative AI, and California reads its fee rule the same way, down to noting that a general AI subscription is typically overhead rather than a billable client cost. Both duties settle in writing once: tell the client what the tool does, and bill the minutes that happened.

An ethical setup, step by step

The duties turn into practice in five steps a small firm can finish in one afternoon. Work through them in order.

1. 01

   Read your own state's guidance first

   State bars differ in emphasis. The three documents in the table above give the pattern, and where your state has spoken, its opinion controls the details. Where it has not, Formal Opinion 512 and the Model Rules are the reference point firms reach for.

   Result: The firm knows which rules and opinions govern it before any tool decision.

2. 02

   Write the firm's AI policy on one page

   California expects supervising lawyers to set clear written policies for AI use. One page is enough for a small firm: what the tool may touch, who reviews, how the work is billed. The prompt below drafts it for a supervising attorney to edit.

   Result: A written policy an attorney has reviewed and adopted.

3. 03

   Set the confidentiality boundary

   Decide which plan the firm uses, confirm its training posture, and write down which document types may go in and which never do. On commercial plans, Anthropic states customer data is not used for training by default; on personal plans that choice is a setting the account holder controls.

   Result: Client data has a defined boundary before any matter touches the tool.

4. 04

   Decide what clients are told

   Consider disclosure in the engagement letter: how the technology is used and what its risks are. Where confidential client information would go into a self-learning generative AI tool, Formal Opinion 512 points to informed consent first.

   Result: A standing answer to the disclosure question, written before a client asks it.

5. 05

   Verify output, then bill the time you spent

   Every citation gets checked against its source, and a licensed attorney reviews every document before anyone relies on it. Hourly bills reflect the drafting and review minutes that happened, never the hours the tool saved.

   Result: Work product that survives a court's scrutiny, and bills that survive a client's.

Step 2 is carried by the prompt below. Copy it into a new chat, fill in the bracketed choices, and hand what comes back to the supervising attorney whose name will be on the policy.

You are drafting an internal one-page AI use policy for a small
law firm. A supervising attorney will review and edit everything
you produce. Mark the top of the document DRAFT FOR ATTORNEY
REVIEW.

Firm: [FIRM NAME]
Practice area: [PRACTICE AREA]
Tool and plan: [TOOL AND PLAN, FOR EXAMPLE: CLAUDE ON A TEAM PLAN]

Write five short sections, each with two or three plain rules:
1. Competence: who owns the tool at the firm, and how its limits
   are learned and kept current.
2. Confidentiality: which document types may go into the tool,
   which never do, and which settings must stay on or off.
3. Verification: who checks citations against sources, and the
   rule that an attorney reviews every document before use.
4. Client communication: what the engagement letter says about
   AI use, and when informed consent is asked first.
5. Billing: time spent is billable, time saved is not.

Rules: plain English, no legal citations, one page at most. Where
a firm decision is missing, write [DECIDE: WHAT THE FIRM MUST
CHOOSE] instead of inventing it.

From there, the safest first matter is a fictional one. The safety page walks that first test end to end, fictional client files included: Is Claude safe for lawyers? To watch a build recorded start to finish, or to have one set up for your firm, the two cards below are the places to go.