Can lawyers use Claude without getting sanctioned?

The specific failure that gets lawyers sanctioned

A language model generates the most plausible next words. It does not look anything up unless you connect it to a source and tell it to. That means a fabricated citation comes out looking exactly like a real one: a normal-sounding case name, a plausible reporter volume, a page number, a jurisdiction, a pin cite, even a quotation and internal citations. Nothing in the output signals which parts are real. The model sounds equally confident either way, and confidence is not evidence.

This is not hypothetical. In Mata v. Avianca, Inc., Judge P. Kevin Castel of the Southern District of New York issued an opinion and order on sanctions on June 22, 2023, after an opposition brief cited six opinions that did not exist: Varghese, Shaboon, Petersen, Martinez, Durden, and Miller. The court acted under Rule 11 and, alternatively, its inherent authority.

Read the actual order rather than the coverage, because the non-monetary part is the part that should get your attention.

• ·A $5,000 penalty imposed jointly and severally on the lawyers who filed it and their firm, payable into the court registry within 14 days
• ·A letter to the client, attaching the sanctions opinion, the hearing transcript, and the affirmation containing the fake opinions
• ·A letter to each real judge who had been falsely named as the author of one of the fabricated opinions, attaching the fake opinion attributed to them
• ·Copies of every one of those letters filed on the public docket

This is not only a ChatGPT problem, and Claude is not exempt

It would be convenient to say this was one careless lawyer in 2023 with one early tool. It was not. Damien Charlotin maintains a database of court decisions worldwide in which a tribunal found that a party relied on AI-generated content that turned out to be fabricated. It listed more than 1,900 decisions as of August 2026, and it grows most weeks. The parties involved include practicing lawyers, self-represented litigants, expert witnesses, and in at least one instance a judge.

Claude appears in that database too. The tools named in a 2026 District of Connecticut matter included Claude alongside two others, in a filing with quotes that did not match the authorities cited, and the result was a $3,500 monetary sanction plus a bar referral. Anyone telling you their preferred model is the safe one is selling you something. Every general-purpose model can produce a citation that is not real, and the duty to verify does not shift to the vendor.

What ABA Formal Opinion 512 actually says

On July 29, 2024, the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, its first formal opinion on generative AI in the practice of law. It does not create new rules and it does not ban anything. It identifies which existing Model Rules do the work, and it says lawyers and firms using these tools must fully consider their applicable ethical obligations.

Four duties carry most of the weight, and none of them are exotic. What is new is only the tool, not the obligation.

The opinion also expects you to keep learning as the tools change, and the ABA said plainly that its committee and state and local bar ethics committees will likely keep issuing updated guidance as specific tools develop. Treat any guidance you read, including this page, as a snapshot rather than a settled answer.

• ·Competence, Model Rule 1.1: you must provide competent representation and understand the benefits and risks of the technology you use. You do not have to be an AI expert. You do have to understand what your specific tool can and cannot do, and you cannot substitute its output for your own judgment.
• ·Confidentiality, Model Rule 1.6: you must keep information relating to a representation confidential regardless of its source, unless the client gives informed consent. Before client information goes into any tool, you need to know how that tool handles it.
• ·Communication, Model Rule 1.4: you must reasonably consult with the client about the means used to pursue their objectives, which can mean telling them you are using these tools.
• ·Fees, Model Rule 1.5: fees and expenses must be reasonable. If the tool makes you faster, an hourly client pays for your actual time, not the time it used to take. In most circumstances you cannot bill a client for the hours you spend learning how to use the tool at all.

The division of labor that holds up

The useful mental model is not risky versus safe software. It is closed-book versus open-book. Asking the model to recall the law from memory is closed-book, and that is where fabrication lives. Handing it the material and asking it to work on what you handed it is open-book, and that is where it earns its keep.

Work that is genuinely safe, because you supplied the source and you can check the output against it: drafting from your own outline or precedent, summarizing a contract, deposition transcript, or discovery production you uploaded, turning your own analysis into a plain-English client letter, drafting routine correspondence and intake follow-ups, generating deposition or interview questions from documents you provided, tightening prose you already wrote, and building the first version of a timeline or chronology from records you gave it.

Work that is not safe, no matter how the answer reads: asking what the law is in your jurisdiction, asking for cases that support a position, asking it to confirm that a case it gave you is real, quoting language it attributes to a case you have not opened, or relying on a summary of an authority you never pulled. Asking the model to check its own citations does not work, because the same process that invented the case can invent the confirmation.

One rule closes the gap. Every citation, quote, holding, and statutory reference gets pulled and read in the firm's own research service, or in the actual slip opinion, by a person, before it goes into anything filed or sent. If it is not in there, it does not go in. That single habit is what separates useful drafting help from the docket entry you will be explaining for years.

Confidentiality, disclosure, and the rules that vary

Anthropic's own Usage Policy is more restrictive here than most lawyers expect. It classifies legal interpretation, legal guidance, and decisions with legal implications as a High-Risk Use Case. For those uses, Anthropic requires a qualified professional to review the content or decision before it is disseminated or finalized, and requires you to disclose to individuals and consumers that AI helped produce your advice or recommendations, at minimum at the start of each session. That policy is versioned and dated, so check the current text rather than trusting a summary.

On client data, the answer depends on which plan you are on, and that distinction matters for Rule 1.6. Anthropic states that by default it does not use inputs or outputs from its commercial products, such as Claude for Work and the Anthropic API, to train its models. For consumer plans, whether your chats are used to improve the models depends on your own privacy setting, plus safety review and any program you opt into. If client information is involved, verify the setting yourself in the account you are actually using, and prefer a business plan where the commercial terms apply.

This page is not legal advice, and it is not an ethics opinion. Formal Opinion 512 interprets the ABA Model Rules, which are a model. Your obligations come from your own state's rules and your own bar's guidance, and those genuinely differ: some states have issued formal advisory opinions, some have published practical guidance, and some have said nothing yet. Individual judges add another layer, with standing orders and local rules that require disclosure or certification of AI use in filings. Before you set firm policy, read your state bar's current guidance, check the standing orders of the courts you appear in, and call your malpractice carrier.

What people ask next

• Can I put client data into Claude?
• Is Claude HIPAA compliant?
• What is Claude for Small Business, and do I need it?
• Can I control what my staff put into Claude?

Worked examples