AI for Attorneys: How to Use It Without Breaking Confidentiality.

The question is not whether lawyers can use AI. They already are. Thirty percent of respondents reported using AI tools in the 2024 ABA Legal Technology Survey, up from 11 percent the year before. The real question is whether you can use it without breaching client confidentiality or drawing a sanction, and the honest answer is yes, if you set the rules first. The lawyers who got into trouble did not get there because they used AI. They got there because they pasted privileged facts into a tool that trains on them, or filed a brief full of cases the AI invented and never checked. Both are confidentiality and competence failures, not technology failures, and both are entirely avoidable.

What the ethics rules actually require

Start with the rules, because they are clearer than the noise around them suggests. The core duties run through the ABA Model Rules of Professional Conduct, which most states track in some form: Rule 1.6 on confidentiality, Rule 1.1 and its comment 8 on the duty of technological competence, Rule 1.4 on communicating with clients, and Rules 5.1 and 5.3 on supervising lawyers and vendors. None of them bans AI. They require you to understand the tool and protect the client.

In July 2024 the ABA issued Formal Opinion 512, its first ethics guidance on generative AI. It is worth knowing what it actually says. Where using an AI tool would disclose a client's confidential information to a third party in a way that would otherwise violate Rule 1.6, you generally need the client's informed consent, and boilerplate language buried in an engagement letter is not enough. Whether a given tool triggers that duty turns on the vendor relationship and how the data is handled, not simply on whether the tool learns from inputs. Informed consent means you actually explain the risk. You must be competent in the specific tool, understand its limits, and independently verify its output. Managing lawyers should adopt reasonable AI policies and supervise staff and vendors. And you ordinarily should not bill a client for the time you spend learning AI in general, or bill saved time as though the work took longer.

States have gone further. California issued practical guidance in November 2023, Florida approved Ethics Opinion 24-1 in early 2024, Texas issued Opinion 705 in 2025, and New York's bar associations have published detailed reports. The thread running through all of them is the same: do not put confidential client information into a tool that lacks adequate protection, and analyze, through the lens of Rule 1.6, whether a tool that would disclose that information to a third party calls for the client's consent. This guidance is interpretive and varies by jurisdiction, and nothing here is legal advice, so check your own state's rules. But the direction is consistent everywhere.

The real risk: consumer tools train on what you type

This is the part that most often goes wrong. The default settings on the consumer versions of these tools are not built for privileged data. Consumer ChatGPT on the Free and Plus tiers uses your conversations to improve the models unless you turn it off, and even then retains content for a period for abuse monitoring. Anthropic changed its consumer policy in 2025: Claude Free, Pro, and Max accounts now train on your chats unless you opt out, and if you allow it, conversations can be retained for up to five years. That reverses the old reassurance that Claude never trains on your inputs, which means the assumption many lawyers are still operating on is out of date.

This is exactly the scenario Formal Opinion 512 had in mind when it tied the consent requirement to self-learning tools. If a tool can retain and learn from a client's confidential facts, those facts can be stored, reviewed by a human, or surface later, and that is the worst case for privileged material. The fix is not to avoid AI. It is to use the right tier. Enterprise, Team, API, and "for Work" versions, along with vetted legal tools like CoCounsel, Harvey, and Lexis+ AI, contractually exclude your prompts and files from training and let you control retention. For anything touching a client matter, that is the defensible choice, and the contract terms are worth reading rather than assuming.

The cases where lawyers got sanctioned

The cautionary tales are real, and they all rhyme. In Mata v. Avianca in 2023, a federal judge in New York sanctioned two attorneys and their firm 5,000 dollars for a brief built on cases ChatGPT had invented. That was the first. It was not the last. In Wadsworth v. Walmart in 2025, attorneys at a large plaintiffs' firm were sanctioned for eight nonexistent cases generated by the firm's own in-house AI platform, with one lawyer's pro hac vice admission revoked, proving that an internal tool still needs human verification. In Coomer v. Lindell that same year, lawyers were fined for a filing riddled with fake citations produced across a mix of AI tools. Courts in Oregon and California added fines of 15,500 and 10,000 dollars. A 2025 case introduced a new twist: lawyers dinged for failing to catch the other side's fabricated citations.

The scale is no longer trivial. A widely tracked database has logged more than 700 court decisions worldwide involving AI-hallucinated content, with the large majority in 2025, and the penalties are climbing from warnings to four-figure fines to disqualifications and bar referrals. Read those cases and the pattern is unmistakable. Not one lawyer was punished for using AI. Every one was punished for failing to check what it produced. AI drafts. You verify. You sign. That order never changes.

The confidentiality-first setup

Put the rules and the risks together and a practical setup falls out. None of it requires you to be technical. It requires you to decide a few things on purpose and write them down.

• Keep privileged facts out of consumer tools. Do not paste client identifiers or confidential facts into a free or consumer tool that may train on them. Use general consumer tools only for non-confidential, general-knowledge work.
• Use the right tier for client work. Move anything touching a matter to an enterprise, API, or "for Work" tier, or a vetted legal tool, that contractually does not train on your data. Turn training off and tighten retention.
• Anonymize and redact. Strip names, matter numbers, and identifying facts before prompting wherever the task allows it.
• Do vendor due diligence. Before you trust a tool, ask where the data is stored, who can access it, whether inputs train the model, how long data is retained, and what security it carries.
• Get informed consent where required. Where a third-party tool would see confidential information, obtain and document the client's informed consent, and explain the actual risk rather than relying on boilerplate.
• Verify every output. Independently confirm every case citation, quotation, and statement of law before it goes into a filing or to a client. For a citation, that means pulling the case, reading the pages that matter, and confirming the holding, not skimming a headnote. This one discipline would have prevented every sanction above.
• Supervise and train. Adopt a written firm AI policy and train your lawyers, staff, and any vetted vendors to it.
• Mind the fees and the standing orders. Do not bill clients for general AI learning or bill saved time as if it took longer, and check whether the judges you appear before require disclosure or certification of AI use.

The rules do not ban AI. They ban using it carelessly with a client's secrets. Set the boundary first, and the tool becomes an asset instead of a malpractice risk.

Where The Leveraged Attorney course fits

This is exactly what we build in the AI for Lawyers course. Not a lecture on the technology, but the confidentiality-first operating setup a practicing attorney needs: the data boundaries, which tier to use for which task, how to anonymize and get consent, the verification protocol that keeps fabricated citations out of your filings, line-by-line firm AI policy language, a vendor diligence checklist, and client consent language you can adapt. It is built so you can put AI to work on real matters with confidence instead of fear, and stay on the right side of the rules while you do it.

The difference between an AI asset and a malpractice risk is the operational boundary you set before you use the tool, not after. Installing that boundary, in language you can actually adopt, is what the AI for Lawyers course is built to do. This briefing is general information and not legal advice; confirm the rules in your own jurisdiction before you rely on any of it.