Can Lawyers Use AI for Legal Research? Ethics, Rules, and Best Practices

Short answer: Yes. No state bar and not the American Bar Association prohibits lawyers from using artificial intelligence for legal research. What every authority does require is this: the AI is your assistant, not your judgment. You remain fully responsible for the accuracy of everything you file and advise. Use AI to find and frame authority faster, then verify every citation against the actual source before it leaves your desk.

That's the rule in one sentence. The rest of this guide explains where it comes from, what each ethical duty actually demands, and how to use AI for legal research without ending up as the next cautionary headline.

This article is for general information and is not legal advice. Ethics rules vary by jurisdiction; consult your state's rules of professional conduct and your bar's guidance before relying on any AI tool in a client matter.

Is it ethical for lawyers to use AI? What the rules say

There is no separate "AI rule" for lawyers. Instead, the existing Rules of Professional Conduct apply to AI the same way they apply to a junior associate, a contract attorney, or a legal database. The American Bar Association made this explicit in ABA Formal Opinion 512{:target="_blank" rel="noopener noreferrer"}, "Generative Artificial Intelligence Tools," issued July 29, 2024 — the ABA's first formal ethics opinion on generative AI.

The opinion's core message is that generative AI is permissible, but it does not dilute any of a lawyer's existing obligations. As the ABA put it, lawyers using these tools must "fully consider their applicable ethical obligations," which the opinion grounds in a familiar set of Model Rules: competence, confidentiality, communication, candor to the tribunal, supervision, and reasonable fees.

In other words: the question is not whether you may use AI. It is how you use it so that you still satisfy the duties you already owe your clients and the court.

The six ethics duties that govern AI legal research

ABA Opinion 512 and the state guidance that followed cluster around six duties. Here is what each one requires when AI enters your research workflow.

1. Competence (Model Rule 1.1)

Model Rule 1.1{:target="_blank" rel="noopener noreferrer"} requires "the legal knowledge, skill, thoroughness and preparation reasonably necessary for the representation." Comment 8 to that rule extends competence to technology: a lawyer should keep abreast of "the benefits and risks associated with relevant technology." A majority of U.S. states have formally adopted some version of this duty of technological competence.

Applied to AI, competence has two edges. You must understand a tool well enough to know what it is good at (surfacing relevant authority, summarizing long records) and where it fails (fabricating citations, missing controlling law, going stale). And you must remain competent in the underlying law, because you — not the model — are the one certifying the work.

2. Confidentiality (Model Rule 1.6)

Model Rule 1.6 requires a lawyer to protect "information relating to the representation," regardless of source, unless the client gives informed consent. Many consumer AI tools retain user inputs, share them with third parties, or use them to train future models. Entering privileged facts or client documents into such a tool can breach confidentiality before you ever hit "file."

Before putting anything client-related into an AI system, ABA Opinion 512 and state guidance direct lawyers to understand the tool's data retention, data sharing, and model-training practices — and, where the risk warrants it, to obtain the client's informed consent. Enterprise or legal-specific tools that contractually exclude your data from training are materially safer than free public chatbots.

3. Communication (Model Rule 1.4)

Model Rule 1.4 obligates lawyers to keep clients reasonably informed and to "reasonably consult" about the means used to pursue their objectives. Whether you must affirmatively tell a client you used AI depends on materiality: if AI use implicates the client's confidential information, or significantly shapes the strategy or work product, disclosure is generally warranted. Routine, low-stakes use that touches no client data usually does not require a specific heads-up — but client engagement letters and a growing number of judicial standing orders increasingly ask for it anyway.

4. Candor to the tribunal and the duty to verify (Model Rule 3.3 and FRCP 11)

This is where most lawyers get into trouble. Model Rule 3.3{:target="_blank" rel="noopener noreferrer"} forbids knowingly making a false statement of law or fact to a court, and Rule 11 of the Federal Rules of Civil Procedure imposes an independent gatekeeping duty: by signing a filing, you certify that its legal contentions are warranted by existing law.

AI tools do not satisfy that duty for you. A model can produce a citation that looks perfect — correct reporter, plausible parties, confident holding — and is entirely invented. The lawyer's obligation is to read and confirm every authority against the primary source before relying on it. (For the mechanics, see our guides on how to research case law and how to validate a citation.)

5. Reasonable fees (Model Rule 1.5)

Model Rule 1.5 requires fees and expenses to be reasonable. ABA Opinion 512 gives a concrete example: if you spend 15 minutes feeding facts into an AI drafting tool, you may bill that 15 minutes plus the time you spend reviewing the output for accuracy. What you generally may not do is bill the client for the hours you spent learning how to use the tool, or bill as though AI-accelerated work took the full pre-AI time. Efficiency gains belong to the client, not the timesheet.

6. Supervision (Model Rules 5.1 and 5.3)

Partners and supervising lawyers must make reasonable efforts to ensure that everyone in the firm — and every tool the firm uses — operates consistently with the ethics rules. ABA Opinion 512 treats AI use as something that requires firm-level oversight: clear policies on which tools are approved, what data may be entered, and how output must be verified. Associates and paralegals using AI need the same supervision as any other delegated work.

The cautionary tale: Mata v. Avianca

The risk is not theoretical. In Mata v. Avianca, Inc.{:target="_blank" rel="noopener noreferrer"}, 678 F. Supp. 3d 443 (S.D.N.Y. 2023), two attorneys submitted a brief opposing a motion to dismiss that cited at least six non-existent judicial opinions — complete with fabricated quotations and internal citations — generated by ChatGPT. When opposing counsel and the court could not locate the cases, the attorneys doubled down, providing fake "copies" of the decisions rather than coming clean.

Judge P. Kevin Castel found that the attorneys acted in "subjective bad faith" and imposed a $5,000 sanction under Rule 11, jointly and severally, on the two lawyers and their firm. Critically, the court was clear that the problem was not the tool. As Judge Castel wrote, "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance." The sanction was for abandoning the gatekeeping duty — for filing unverified authorities and then standing behind them.

The lesson maps directly onto Opinion 512: AI is allowed; unverified AI output is not. (We dug into why these tools invent citations in the first place in why AI legal citations are so often wrong.)

What state bars say

The ABA opinion is influential but not binding — your state's rules control. A wave of state and local bar guidance, most of it issued in 2023 and 2024, lands on the same principle: AI is permitted, with the lawyer fully responsible for the output.

The details differ at the margins, but no jurisdiction bans AI legal research, and none lets the lawyer outsource responsibility for the result. Always check your own state's current rules and any opinions issued since these.

Do you have to tell clients or courts you used AI?

Clients: It depends on materiality. Disclosure is generally required when AI use touches the client's confidential information or meaningfully shapes the representation; it is generally not required for routine, behind-the-scenes use that implicates neither. Many firms now address AI use directly in their engagement letters, which is the cleanest way to set expectations.

Courts: Watch for standing orders. A number of federal and state judges have issued standing orders requiring attorneys to disclose, or certify the verification of, any AI-assisted filing. These vary by judge and change frequently, so check the specific judge's rules in every matter — a firm-wide assumption is not enough.

Best practices: using AI for legal research without risking sanctions

A practical checklist that keeps you on the right side of every duty above:

1. Treat AI output as a lead, never as authority. Pull and read the actual opinion or statute before you cite it. If the source does not exist or does not say what the tool claims, it does not go in the brief.
2. Use tools that link to primary sources. Prefer legal-specific tools that tie every assertion to a real, verifiable citation you can open and confirm. The faster you can get from claim to source, the safer your workflow.
3. Protect client data. Know your tool's data retention and training policies. Do not enter confidential information into systems that train on inputs or lack adequate security, and obtain informed consent where the risk warrants it.
4. Confirm jurisdiction and currency. Verify that authority is controlling in your forum and still good law. AI is prone to surfacing overruled, reversed, or out-of-jurisdiction cases.
5. Check the judge's standing orders and your engagement letter. Disclose AI use where required, and set client expectations in writing up front.
6. Bill honestly. Charge for time actually spent operating and reviewing the tool, not for learning it, and pass efficiency gains to the client.
7. Adopt a firm AI policy. Define approved tools, permissible data, and a mandatory verification step. Supervise AI-assisted work like any other delegated task.

Done this way, AI does what it is genuinely good at — compressing hours of research into minutes — while you retain the judgment and verification that the rules require. This is the model CaseRead is built around: every case surfaced is real and linked directly to the opinion, so verification is a click, not a leap of faith. The tool finds the authority; you confirm it and own it.

Frequently asked questions

Can lawyers ethically use AI for legal research? Yes. Neither the ABA nor any state bar prohibits it. ABA Formal Opinion 512 confirms that generative AI is permissible under the existing Model Rules, provided the lawyer meets the duties of competence, confidentiality, communication, candor, and reasonable fees — and independently verifies every output.

Do lawyers have to tell clients they used AI? Sometimes. Disclosure is required under Model Rule 1.4 when AI use is material — for example, when client confidential information is entered or when AI output significantly shapes the work. Routine use that touches no client data generally does not require specific disclosure, though engagement letters and some court standing orders may.

Can a lawyer be sanctioned for using AI? Not for using it, but for failing to verify it. In Mata v. Avianca (S.D.N.Y. 2023), the court imposed a $5,000 Rule 11 sanction on attorneys who filed fake ChatGPT-generated citations. The sanction was for submitting unverified, non-existent authorities.

Can lawyers bill clients for time spent using AI? Yes, for reasonable time actually spent operating the tool and reviewing its output. Under Model Rule 1.5 and ABA Opinion 512, lawyers generally may not bill clients for time spent learning a new tool, or charge as if AI-accelerated work took the full pre-AI time.

Is entering client information into an AI tool a confidentiality violation? It can be. Model Rule 1.6 requires protecting information relating to the representation. Before inputting client data, understand the tool's retention, sharing, and training practices, and obtain informed consent where warranted. Tools that train on inputs or lack security pose the greatest risk.