New Mexico AI Ethics Rules for Lawyers

The official instruments

A formal opinion from a smaller bar that punches above its weight

New Mexico is not a state most national AI surveys lead with, and that is part of what makes its instrument worth reading carefully. Formal Ethics Advisory Opinion 2024-004, dated September 24, 2024, is a numbered formal opinion of the State Bar of New Mexico, not a committee FAQ, a website note, or an informal advisory. A smaller bar producing a full formal opinion is a deliberate choice, and the content rewards the attention: 2024-004 covers the standard duties competently, then goes further than many larger states on two points its peers tended to gloss, verification discipline and self-learning conflicts. For a New Mexico lawyer the citation value is straightforward, and for an out-of-state firm with New Mexico matters the opinion is specific enough to build a compliance posture against without guesswork.

The foundation is competence as tool-literacy. Before using any generative AI tool, a New Mexico lawyer must understand how that specific tool handles data, what it does with inputs, whether it retains or trains on them, and where its outputs come from. This is not a demand for engineering knowledge; it is a demand that the lawyer not treat the tool as a black box whose behavior is somebody else’s problem. The opinion then ties that understanding directly to confidentiality: a lawyer must never input confidential or identity-revealing information into tools lacking protections, and must anonymize client details as a matter of course. The phrase “identity-revealing” is doing real work, because it reaches data that is not facially privileged but still lets a third party reconstruct who the client is. Anonymization is therefore the default discipline, not an optional precaution for high-stakes matters only.

Verification New Mexico’s way: outside the model, every time

Where 2024-004 is sharpest is verification. The opinion requires lawyers to independently verify all research, citations, and analysis through traditional databases before filing, and it is explicit that re-asking the same AI does not count. That second clause is the part many practitioners get wrong. The intuitive but mistaken instinct, when an AI produces a citation, is to ask the same system whether the citation is real, which simply invites the model to defend its own fabrication. New Mexico forecloses that shortcut by name: the confirming source has to be an authoritative database outside the model, the same place a lawyer would have checked authority before generative tools existed. This is the strongest possible statement of the anti-hallucination duty short of a court rule, and it converts the sanctions caselaw that prompted these opinions into a concrete workflow requirement. A New Mexico filing that rests on an unverified AI citation is not a close call.

The opinion’s second distinctive contribution is its treatment of self-learning tools and conflicts. Generative AI systems that learn from user inputs can create former-client conflict problems, because information absorbed during one representation can surface, directly or in effect, during another. Most early instruments framed AI risk purely as a confidentiality-in, garbage-out problem; New Mexico recognized that a learning tool is also a vector for conflicts of interest and for the leakage of one client’s information into another client’s matter. The practical consequence is a strong preference for closed tools that do not train on inputs, or, where an open tool is used, rigorous input discipline so that nothing capable of resurfacing ever enters the system. This is also why the confidentiality posture New Mexico adopts is best implemented through an approved-tools register that records each tool’s training behavior, not just its security.

What New Mexico lawyers should do now

The remaining duties round out a complete program. New Mexico requires firm AI policies and training, locating supervision at the firm level, and it lists specific triggers for discussing AI use with clients rather than imposing a blanket disclosure rule, so communication is a defined-circumstances duty in line with the national pattern. Putting it together, a New Mexico firm should: maintain an approved-tools register that captures each tool’s data handling and, critically, whether it trains on inputs; default to anonymization for every input, treating identity-revealing data as off-limits to unprotected tools; enforce a verification workflow that checks citations and analysis in a traditional database and never accepts the model’s own confirmation; screen for the conflicts the opinion flags by preferring closed tools; and adopt a written, acknowledged AI policy with training. The policy template implements each element, and the tracker with its changelog will move New Mexico’s entry the week anything changes it.

One layer 2024-004 does not address is the courtroom. New Mexico has no statewide court rule on generative AI that this tracker has verified, so the obligations that exist at the filing stage are judge-by-judge standing orders rather than a uniform state command. A New Mexico lawyer relying on AI-assisted work should confirm the assigned judge’s standing orders before every filing, because a disclosure or certification requirement can exist at the chambers level even though the state bar’s opinion stops at the verification duty. The court orders guide covers that per-filing check and the kinds of orders to watch for.