Most of the commentary on ABA Formal Opinion 512 has read it as a caution sign. I build the tools it worries about, and I read it as a specification.
The opinion, which the ABA’s Standing Committee on Ethics and Professional Responsibility issued in July 2024 as its first comprehensive guidance on generative AI, does not tell lawyers to avoid the technology. It tells us something harder. Model Rule 1.1 competence, it says, requires “a reasonable understanding of the capabilities and limitations” of any generative AI tool we use. Not expertise. Not a computer science degree. An understanding sufficient to know what the tool can be trusted to do and where it will fail. And it reminds us that the duties we already carry (confidentiality under Rule 1.6, meritorious contentions and candor under Rules 3.1 and 3.3, supervision under Rules 5.1 and 5.3, reasonable fees under Rule 1.5) do not bend because the work product came out of a text box.
What most discussions of 512 miss is that “reasonable understanding” is not a state of mind. It is an architecture question.
The limitation that matters
A large language model, left to its own devices, generates legal citations the same way it generates everything else, by producing plausible text. It will report Smith v. Jones, 512 So. 2d 100 (Fla. 3d DCA 1987), with perfect Bluebook form, complete confidence, and no obligation to the actual Southern Reporter. NIST’s generative-AI risk profile calls this “confabulation,” which is more precise than “hallucination” because the system is not seeing things that aren’t there. It is confidently filling gaps with fabrication, the way a confabulating patient does.
The sanctions cases keep coming. Since Mata v. Avianca made the genre famous, courts have sanctioned lawyers in hundreds of reported incidents, and I do not believe it is because the profession is uniquely lazy (although I’m willing to consider arguments to the contrary). It is because lawyers keep treating a limitation of the architecture as a failure of diligence. You cannot prompt your way out of confabulation. You cannot “be careful” your way out of it, any more than you can proofread your way out of a citation you have no reason to doubt. The tool’s failure mode is engineered to be invisible to the very review 512 tells you to perform.
Which means the review has to be engineered too.
What “reasonable understanding” looks like in software
I spent this year building a legal research platform, and the single design decision I am most certain about is that every quotation the system attributes to a court gets checked, verbatim, against the text of the opinion itself, and every citation gets resolved against a citator before it is allowed to look authoritative. Not because a rule required it, but because I know what the generation layer does when nobody is checking. The verification is not a feature; it is the load-bearing wall. That, I would argue, is what Opinion 512’s competence duty actually cashes out to for the practicing bar.
Start with which layer you are trusting. Retrieval, which finds real documents and quotes them, and generation, which produces fluent text, are different operations with different failure modes. A tool that grounds its output in retrieved authority and shows you the source is categorically different from one that answers from its training weights. Ask which one you are holding before you rely on it.
Then treat verification as the workflow, not an aspiration. If your tool does not verify citations against primary sources, then you are the verification layer, every time, with no exceptions for the citations that look most plausible. Plausible-sounding is precisely where the machine excels.
Know where the data goes. 512’s confidentiality analysis turns on whether client information leaves your control and whether the tool learns from it. That is not a question the marketing page can answer; it is a question about architecture, retention, and training defaults. If you cannot answer it, Rule 1.6’s informed-consent analysis answers itself.
And bill for the work, not the theater. 512 is blunt on fees. Efficiency gains belong to the client. If the machine drafted it in four minutes, Rule 1.5 does not let you bill the four hours it used to take. And 512 says the corollary out loud. Your learning curve is not billable either.
The Florida footnote
Florida lawyers did not have to wait for the ABA. The Florida Bar’s Ethics Opinion 24-1, issued in January 2024, walked the same ground six months earlier, covering confidentiality first, meaningful lawyer oversight, fee integrity, and a distinctly Floridian word of caution about advertising claims. The two opinions rhyme, and together they make the compliance picture clear enough that no Florida practitioner should be surprised by it. The duty was never “don’t use the machine.” It was always “remain the lawyer.”
Remaining the lawyer
The uncomfortable honesty at the bottom of Opinion 512 is that it assigns responsibility to the party least equipped to inspect the failure, the practitioner staring at output that looks exactly like competent work. That is an unstable place to leave the profession, and I don’t think we should stay there. The stable equilibrium is tools that make the verification cheap, visible, and default, so that checking a citation costs one click instead of one trip to Westlaw and a fabricated authority gets flagged before it gets filed.
Until your tools do that, assume they don’t. Verify like the sanction is real, because it is. Understand the machine like the rule requires, because it does. And remember that between the LLM and the lawyer, there is exactly one licensed professional, and it isn’t the LLM.