Lawyers and regulated professions: the duties did not move

Professional duties are the rare part of this field that did not change. What changed is how cheaply they can now be breached: a fabricated citation costs one prompt, a confidentiality breach costs one paste. The two documents this page reads (a sanctions order and an ethics opinion) map exactly those two failure routes.

What actually drew the sanction

Mata v. Avianca is the case everyone half-knows: a brief citing six judicial opinions that did not exist, generated by a chatbot. The half nobody quotes is the court's opening position: "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance". The tool was never the offence.

The offence was what came after. "existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings", and the lawyers "abandoned their responsibilities when they submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT" — and then defended the fakes when challenged. The order lists the damage plainly: "Many harms flow from the submission of fake opinions" — the opponent's wasted work, the court's, and real judges' names attached to invented rulings.

Fabricated authority is the known failure mode of these tools, and legal citations are its natural habitat: dense, formulaic, and expensive to check. The gatekeeping duty prices that in: every citation gets opened before it gets filed. Not because a rule about AI says so, but because the old rule about accuracy never left.

The duty structure, as the ethics opinion reads it

The ABA's Formal Opinion 512 (quoted here via the NCBE's Bar Examiner, since the ABA's own site refuses automated access) walks the familiar duties through the new tool. Three carry most of the load.

Competence. The opinion flags "lawyers' uncritical reliance on content created by a GAI tool" as the risk, and refuses a one-size answer on how much checking is enough: the required verification "will necessarily depend on the GAI tool and the specific task that it performs". Brainstorming needs little; anything you could not judge unaided needs the most. The Bar Examiner's summary sentence is the whole page in one line: "GAI tools cannot solely substitute for a lawyer's competent legal work".

Confidentiality. The protected class is "all information relating to the representation of a client, regardless of its source" — broader than privilege, broader than documents. The opinion's sharpest concern is tools that train on input, where entering client information creates "the risk that information relating to one client's representation may be disclosed improperly, even if the tool is used exclusively" within the firm. That makes training and retention settings (Claude, ChatGPT, Gemini) a professional-duty question, and consent real work: "Boilerplate waivers will not suffice as informed consent".

Supervision. "Managerial lawyers must establish clear policies regarding the law firm's permissible use of GAI" — the firm-level version of an acceptable-use policy, with the same corollary: a firm with no policy has a shadow-use policy.

The questions your jurisdiction answers differently

Where the US analysis stops, local law starts, and three questions are worth asking your own regulator rather than this page: whether client consent is needed before any client data enters a third-party tool at all, whether your professional-secrecy statute (in Switzerland, Art. 321 StGB reaches further than most confidentiality rules) treats a cloud AI processor like any other provider, and what your courts have said about disclosure of AI use in filings. The pattern travels; the answers do not.

What goes wrong

The unopened citation. Filed, challenged, defended — the Mata sequence. The failure was never the generation; it was the missing five minutes of checking.

The client file in a consumer account. Confidentiality breached at paste time, on a personal subscription with default settings nobody read (Claude, ChatGPT, Gemini).

The firm ban that produces shadow use. Usage moves to personal devices, outside every control, and the first the firm hears is the incident.

"The enterprise tier handles consent." Procurement settings reduce risk; they are not the client's informed consent, and the opinion is explicit that boilerplate does not manufacture it.

Delegated judgement. Drafting is labour; assessing the argument is the professional act. The deskilling page covers why the second erodes quietly while the first gets delegated.

How to check it worked

Run the Mata drill on your own last AI-assisted document: open every cited authority and confirm it says what the document claims. Then run the confidentiality drill: list which tools touched client information this month, and for each, answer what its retention setting is and whose account it ran under. A firm that can answer both in an hour has a practice; one that cannot has an incident with a date not yet filled in.