State Bar Ethics Opinions on AI Use by Lawyers

There is no ethics rule about artificial intelligence anywhere in the United States. What exists is a set of opinions applying rules written decades earlier — competence, confidentiality, supervision, candour, fees — to a new tool, and the opinions are unusually consistent about which rules do the work.

ABA Formal Opinion 512

The American Bar Association’s Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, Generative Artificial Intelligence Tools, in July 2024. The ABA does not regulate anyone: the Model Rules of Professional Conduct are a template that states adopt, adapt or ignore, and a formal opinion is persuasive rather than binding anywhere. It matters because state committees cite it and because it is the most complete treatment.

Its analysis runs through six rules, and the structure is worth keeping because every state opinion since follows some version of it.

• Rule 1.1, competence. Comment [8] requires a lawyer to keep abreast of the benefits and risks associated with relevant technology. Applied to generative tools this is not a duty to use them, and not a duty to understand transformer architecture; it is a duty to understand enough about a specific tool’s capabilities and limitations to use it responsibly — including that it can produce fluent, confident, false output.
• Rule 1.6, confidentiality, with the reasonable- efforts requirement at 1.6(c).
• Rule 1.4, communication — when the client must be told the tool is being used.
• Rules 5.1 and 5.3, supervision of lawyers and of non-lawyer assistance, which the opinion treats as covering AI tools used by staff.
• Rule 1.5, fees, and the prohibition on unreasonable charges.
• Rules 3.1 and 3.3, meritorious claims and candour to the tribunal, which is where fabricated authority lands.

Confidentiality and informed consent

The most operationally significant conclusion in Opinion 512 concerns inputs. The opinion distinguishes tools that use inputs to further train or improve a model from those that do not, and takes the position that before entering information relating to the representation of a client into a self-learning generative AI tool, a lawyer must obtain the client’s informed consent.

Informed consent under Model Rule 1.0(e) requires the client to be told enough to appreciate the risk, which means an explanation of what happens to the data rather than a line in an engagement letter saying the firm uses technology. That is a demanding standard and it is why the input question dominates law-firm procurement: a tool with a contractual no-training commitment and defined retention changes the analysis, because the risk the client is being asked to accept is different.

Confidentiality also runs sideways. Rule 1.6 covers information relating to the representation, not only privileged information, and it covers other clients’ matters. A retrieval system indexing the firm’s entire document store and answering any lawyer’s question crosses ethical walls that exist to manage conflicts under Rules 1.7 and 1.10. That is a configuration problem with an ethics consequence, and it is not addressed by any vendor’s security certification.

Verification, and the rule it comes from

Every opinion in this area says output must be verified. It is worth being precise about where that duty comes from, because the source determines the standard.

It is not a special AI rule. Rule 3.3(a)(1) prohibits a lawyer from knowingly making a false statement of fact or law to a tribunal, and Rule 3.1 requires a basis in law and fact for claims asserted. In federal practice, Rule 11(b) of the Federal Rules of Civil Procedure certifies that legal contentions are warranted by existing law, and courts have relied on Rule 11 rather than on any AI-specific authority in sanctioning lawyers who filed fabricated citations. The best-known instance is Mata v. Avianca, Inc. in the Southern District of New York, where sanctions were imposed in June 2023 — and the court’s reasoning turned substantially on what the lawyers did after the citations were questioned, not merely on having submitted them. The pattern of later decisions is covered in the page on fabricated citations and sanctions.

The practical standard the opinions describe is that verification means reading the cited authority in a reliable source and confirming it says what the output claims. Asking a model whether its own citations are real is not verification: the same process that produced the citation produces the confirmation. Nor is finding a case with a matching name sufficient, since fabricated citations frequently attach a real case name to an invented holding or a wrong reporter.

Fees and billing

Opinion 512 addresses fees directly and reaches two conclusions that are more concrete than most ethics guidance.

A lawyer billing hourly may charge only for time actually expended. If a tool reduces a task from five hours to one, the client is billed for one; the efficiency belongs to the client, not to the firm. And a lawyer may not bill a client for time spent learning to use a general-purpose tool, which is analogous to the long-standing position on overhead and general education. Charging a client for the cost of a specific tool used on their matter can be permissible if disclosed in advance and reasonable, and the opinion treats that as a disclosure-and-reasonableness question under Rule 1.5 rather than a prohibition.

The related trap is advertising. Rule 7.1 prohibits false or misleading communications about a lawyer’s services, and several state opinions warn specifically against claims that a firm’s use of AI makes it superior to others or produces better outcomes. That is the same exposure that federal regulators pursue against companies generally, discussed in the page on AI marketing claims.

Opinions by issuer and date

The following are published opinions and guidance documents on generative AI from bar authorities, with issuer and date. This list is not exhaustive — new opinions have continued to issue — and each entry should be read in its own text rather than through this summary.

• American Bar Association, Formal Opinion 512, Generative Artificial Intelligence Tools, July 2024. Available through the ABA’s ethics opinions library.
• State Bar of California, Committee on Professional Responsibility and Conduct, Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, approved November 2023 — the first substantial guidance from a US bar authority. Published by the State Bar of California.
• The Florida Bar, Ethics Opinion 24-1, January 2024, addressing confidentiality, oversight of generative AI as analogous to non-lawyer assistance, fees, and advertising claims. Florida Bar ethics opinions.
• Supreme Court of New Jersey, Preliminary Guidelines on the Use of Artificial Intelligence by New Jersey Lawyers, January 2024 — notable for coming from the court rather than a bar committee, published by the New Jersey Courts.
• Pennsylvania Bar Association and Philadelphia Bar Association, Joint Formal Opinion 2024-200, May 2024.
• Kentucky Bar Association, Ethics Opinion E-457, 2024.
• District of Columbia Bar, Ethics Opinion 388, 2024, on attorneys’ use of generative AI in client matters. D.C. Bar ethics opinions.
• New York City Bar Association, Committee on Professional Ethics, formal opinion on generative AI, 2024.
• State Bar of Texas, Professional Ethics Committee, Opinion 705, 2025.

What these opinions do not decide

Three questions come up constantly and are not answered by any of the opinions above.

Whether use must be disclosed to the client. The consensus is that Rule 1.4 requires disclosure in some circumstances — where the client asks, where the tool’s use is material to the representation, where confidential information is being entered — and not as a blanket rule for every use of every tool. Nobody has drawn the line, and the opinions differ in emphasis. Using a model to draft a routine letter is not treated the same as using one to analyse a settlement position, and no test distinguishes them.

Whether privilege survives. The attorney-client privilege is a doctrine of evidence law, not of professional responsibility, and an ethics committee cannot settle it. Disclosure to a third-party vendor is generally analysed under the same framework as disclosure to any agent assisting the lawyer, which usually preserves privilege, but the analysis depends on the vendor’s terms, its rights over the data and its access to content. There is little case law directly on generative AI vendors, and the safe assumption is that the question is open.

Whether courts will require disclosure in filings. A number of federal judges adopted standing orders in 2023 and 2024 requiring certification about generative AI use in filings, and the practice varies widely by judge and district; some courts have declined to adopt any such requirement, reasoning that Rule 11 already covers it. There is no uniform rule, and the operative requirement is whatever the individual judge’s standing order says. Check it for each matter, because it is the one obligation on this page that can differ between two courtrooms in the same building.