Can Lawyers Use AI for Legal Research? Yes, With Conditions

Yes. Lawyers may use AI for legal research. ABA Formal Opinion 512, issued 29 July 2024, confirms that generative AI is compatible with the Model Rules provided the lawyer meets the existing duties of competence, confidentiality, client communication, candor to the tribunal, supervision and reasonable fees. No US jurisdiction bans it. What courts sanction is filing authority nobody read, which is a Rule 11 problem rather than an AI problem.

That distinction sounds like a technicality until you read the sanctions decisions, and then it turns out to be the entire practical lesson. Not one of the 2026 federal decisions penalized an attorney for using a machine. Every one of them penalized an attorney for putting a citation in front of a court without opening it. The tool is not the violation. The unverified filing is.

So the useful question is not whether you are allowed to use AI. It is which kind of AI, for which task, with which verification step in between. This walks through all three, with the current US authority. It is informational research guidance, not legal advice.

Can lawyers use AI for legal research?

Yes, and the ABA said so explicitly. Formal Opinion 512, "Generative Artificial Intelligence Tools", was the first ABA ethics guidance written specifically about generative AI. Its approach is to map the existing Model Rules onto the technology rather than to create a new regime, which is why it reads as a checklist of duties you already have. The opinion is clear that these tools cannot substitute for a lawyer's own competent work, and equally clear that using them is permitted.

State bars have followed with their own guidance and it is not uniform. California, Florida, New York, New Jersey, Texas and others have issued opinions or task force reports, and they differ on emphasis, particularly around disclosure and billing. Read your jurisdiction's opinion alongside the ABA's rather than assuming Formal Opinion 512 settles the question for you.

What does ABA Formal Opinion 512 actually require?

Six duties, in the opinion's own framing:

• Competence, Rule 1.1. Understand the capabilities and limits of the tool you are using, and keep that understanding current as the tool changes. You do not have to be an engineer. You do have to know what it is doing when it produces a citation.
• Confidentiality, Rule 1.6. Know how the tool handles the data you put in it and put safeguards around it. The opinion goes further than most lawyers expect: it says you should generally obtain informed client consent before entering client confidences into a generative AI tool, and that boilerplate consent tucked into an engagement letter does not do the job.
• Communication, Rule 1.4. Tell the client when the use of the tool is material to the representation.
• Candor to the tribunal, Rule 3.3. What you file is your representation to the court, whoever or whatever drafted it.
• Supervision, Rules 5.1 and 5.3. Firm leadership has to put policies and training in place, and the duty extends to non-lawyer assistance, which is the frame the opinion applies to the tools.
• Fees, Rule 1.5. Fees stay reasonable. If a task now takes twenty minutes instead of three hours, you bill twenty minutes, and you do not bill the client for learning the software.

The confidentiality duty is where most firm policies are thinnest. Knowing how a tool handles your input is a procurement question before it is an ethics question, and it comes down to two things you should ask in writing: does the vendor train on your data, and how long is it retained. Firms deploying AI across multiple workflows increasingly put a policy layer that controls what data an AI system can reach in front of the tools rather than relying on each vendor's defaults, which is a sensible answer to a duty that the ABA has placed squarely on you.

Why do lawyers get sanctioned for using AI?

They do not, strictly. They get sanctioned for fake citations. The mechanism is worth understanding because it explains which tools are dangerous and which are not.

A general-purpose language model does not look anything up. It produces text that statistically resembles legal writing, and legal writing contains citations, so it produces things shaped like citations: a plausible party name, a plausible reporter, a plausible volume and page. The output is fluent and formatted correctly and refers to a case that has never existed. Nothing in the model's behavior flags the difference, because from the model's perspective there is no difference.

Damien Charlotin maintains a public database of decisions in which a court found, or clearly implied, that a party relied on hallucinated material. As of its 8 August 2026 update it holds 1,868 cases. The inclusion bar is strict, so that number is a floor rather than an estimate.

What have courts actually done about it in 2026?

Norton Rose Fulbright collected six federal decisions from a ten week window, February to April 2026, and the range of outcomes is the useful part. Two produced no monetary sanction. One produced a five figure penalty per attorney.

• Fletcher v. Experian Information Solutions, Inc., Fifth Circuit, 18 February 2026, 168 F.4th 231: a 2,500 dollar monetary sanction.
• In re: Eric Chibueze Nwaubani, Fourth Circuit, 11 March 2026, 2026 WL 687194: public admonishment.
• Whiting v. City of Athens, Tennessee, Sixth Circuit, 13 March 2026, 170 F.4th 455: the appellees' reasonable fees on appeal, double appellate costs, and 15,000 dollars each in punitive sanctions.
• United States v. Farris, Sixth Circuit, April 2026, 171 F.4th 920: no monetary sanction, but Criminal Justice Act compensation denied.
• Gamez v. County of Fresno, E.D. California, 9 April 2026, 2026 WL 925944: no sanctions imposed.
• Fivehouse v. U.S. Department of Defense, E.D. North Carolina, 27 April 2026, 2026 WL 1278575: public reprimand.

Whiting is the one to read. The Sixth Circuit found that counsel had filed briefs containing more than two dozen fake citations, along with quotations that did not appear in the cases cited and citations that did not support the propositions attached to them, and treated that as misconduct in arguing the appeal. The financial exposure was not the 15,000 dollar penalty. It was the order to pay the other side's appellate fees on top of it.

Read across the six, the pattern that separates a reprimand from a five figure sanction is not the technology and not the number of bad citations. It is what the lawyer did once the problem surfaced. Prompt acknowledgment and correction landed at the reprimand end. Compounding the error with misrepresentation landed at the other. We go deeper into that pattern in our piece on AI hallucinations in legal cases.

Do lawyers have to disclose AI use in court filings?

It depends on your judge, and the answer genuinely varies chambers to chambers. There is no blanket federal rule. Since 2023 a growing number of individual judges have issued standing orders on generative AI, and some jurisdictions have adopted their own procedural rules. They fall into roughly three shapes: no requirement at all, a certification that any AI-assisted citations were verified against an authoritative source, and full disclosure of the fact and extent of AI use.

The practical rule is simple. Check the standing order of the specific judge before every filing, because the requirement lives at the chambers level and it changes. Treat any inherited firm template that says "our courts do not require disclosure" as out of date.

Is AI legal research reliable?

It depends entirely on what the tool searches, and this is the one technical distinction worth carrying with you.

A tool that generates answers from a model's own training has no source to show you, because there is no source. A tool that retrieves from an actual body of case law and then summarizes what it found has a source you can open in one click. Both present the answer in the same confident prose. Only one of them can be checked, and the check takes about thirty seconds.

So the test during a trial period is not "is the answer good." It is "can I open the case, and does the case say what the summary said." Run that on a question you already know the answer to, twice, before you form an opinion about a product. Our comparison of AI legal research tools goes through what each vendor searches, and can ChatGPT do legal research covers why a general assistant fails this specific test.

How should a firm verify AI legal research?

Four steps, in order. None takes long, and together they catch essentially everything in the sanctions record.

1. Confirm the case exists. Open it. Not the summary, the case. If a tool cannot take you to the opinion, the citation is unverified and should not go in a filing. This step alone would have prevented every decision listed above.
2. Confirm the case says what the tool said. Hallucination has a subtler second form: a real case, cited for a proposition it does not support. That is what the Sixth Circuit found alongside the fabricated citations in Whiting, and it is harder to spot because the citation checks out.
3. Confirm it is still good law. A real, correctly characterized case that has been overruled is still a bad citation. This needs a citator. Be aware of the limits: Paul Hellyer's study in Law Library Journal volume 110 reviewed 357 citing relationships and found that Shepard's and KeyCite each missed or mislabeled roughly a third of negative citing relationships, and BCite more than two thirds. Our guide to legal citators covers what each one does and does not catch.
4. Confirm the jurisdiction binds you. A well-reasoned case from the wrong court is persuasive at best. See binding versus persuasive precedent if that line is fuzzy in a particular posture.

Write those four steps into the firm's AI policy as a checklist rather than a principle. Formal Opinion 512 puts the supervision duty on partners and managers, and a checklist is auditable in a way that "exercise professional judgment" is not.

What is the safest way to use AI for legal research?

Use it to find candidates, never to state the law. The AI's job ends when it hands you a list of cases that might be on point. Your job starts when you open them. Used that way it is a very good first-pass researcher, because reading twelve cases to find the three that matter is exactly the work that used to consume an afternoon.

Two habits make this stick. Do not paste a citation from any tool into a brief without having had the opinion open in another tab, and keep the confidential facts out of general consumer assistants entirely until you have satisfied the Rule 1.6 analysis. Neither costs anything. Together they cover the overwhelming majority of the reported failures.

What to take away

Lawyers can use AI for legal research, and the ethical framework for doing so has been settled since July 2024. The risk lives in one specific place: a citation that entered a filing without a human opening the case. The 1,868 decisions in the hallucination database are, almost without exception, that single failure repeated.

The tooling answer follows from the diagnosis. Prefer research tools that retrieve from real case law and show their sources over general assistants that generate text, keep a citator in the workflow, and treat every citation you have not personally opened as unverified. If you are working out which category a given product falls into and what it should cost, our page on legal AI for lawyers lays out the tools task by task with published pricing.

Cases is built for the first step of that workflow. Ask a research question in plain English, get on-point US federal and state precedents with the holding and a citation you can open and read. We do not run a citator and we say so. Informational research, not legal advice, always verify the citation.