Lawyers are asking a simple question and getting a complicated answer. Yes, a law firm can use generative AI. No, it cannot paste confidential client information into a public chatbot and call it fine. The line between those two sits in ABA Formal Opinion 512, and it is clearer than most firms think.
In July 2024 the American Bar Association's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, the first comprehensive guidance on lawyers' use of generative AI. State bars have followed with their own opinions. The practical conclusions line up, and they come down to four duties.
Before anything else
This is a practical overview, not legal advice, and your state's rules control. Read the opinion itself and your bar's guidance before changing how your firm works. The primary source is ABA Formal Opinion 512, as summarized by the ABA Business Law Section.
Duty 1: Confidentiality, and the self-learning problem
Model Rule 1.6 requires lawyers to protect information relating to a representation. Opinion 512 applies that to generative AI directly: lawyers are responsible for knowing how a tool uses data and for putting safeguards in place so that what they feed it stays secure and is not disclosed to third parties.
The specific risk the opinion names is the self-learning tool. If a product trains on what users enter, one client's information can surface in another matter, inside your firm or outside it. The opinion also warns that several lawyers using the same tool may inadvertently disclose client information to one another through it. A consumer chatbot that uses conversations to improve its models is exactly that kind of tool unless you have turned the setting off and can prove it.
Duty 2: Informed consent, and why boilerplate fails
Before inputting information relating to a representation into a self-learning tool, the opinion says lawyers must assess the disclosure risk and obtain the client's informed consent. It is specific about what does not count: boilerplate consent included in engagement letters will not be adequate. The client has to understand what tool, what data, and what risk, and agree to it.
For a firm that wants to use AI across hundreds of matters, that is a real operational burden. It is also the point. The opinion is pushing firms toward tools where the question does not arise, because nothing leaves.
Duty 3: Competence
Rule 1.1. The opinion says lawyers need not become GAI experts, but they must have a reasonable understanding of the tool's capabilities and limitations, or draw on someone who does, and keep that understanding current as the technology changes. In practice that means someone at the firm can answer, in plain language, where the data goes, whether the vendor trains on it, and what the retention policy is.
Duty 4: Fees
Lawyers may not bill clients for time spent learning a technology used for client matters generally. They may bill for learning a specific tool if the client requested it. AI costs can be treated as overhead, or a portion of an expensive tool can be passed through, or per-use charges can apply where appropriate, but any such approach must be explained in advance with the client's informed consent.
What this means for a public chatbot
| Question the opinion asks | Free or consumer chatbot | Enterprise cloud tier | On-site model |
|---|---|---|---|
| Does the vendor train on what you enter? | Often yes by default, opt-out where offered | Usually no, by contract | No vendor; the model runs on your hardware |
| Who else can see it? | The vendor, under its terms | The vendor, under a negotiated agreement | Only your firm |
| Is informed consent needed before inputting client information? | Yes, per matter, not boilerplate | Assess the terms; likely still disclose | No third party, so no disclosure to consent to |
| Can you show a client where the data went? | Only by reading the vendor's policy | Yes, through the contract and logs | Yes, through logs on your own systems |
The consumer tier fails Duty 1 and makes Duty 2 expensive. The enterprise tier can pass, at enterprise prices and with a terms-of-service review on every renewal. The on-site model removes the third party, which removes the disclosure, which removes the per-matter consent conversation.
What a law firm deployment looks like
A dedicated inference machine on the firm's own network, open-weight models tuned for the firm's document types, and agents wired into the document management and practice management systems the firm already runs. Work the firm has moved first, in our experience:
- Intake summaries from client-supplied documents, with every extracted fact traced to a page
- Contract and filing review that flags missing clauses and dates against a checklist the partners wrote
- Transcript to structured plan: a client call becomes decisions, owners and next steps
- Read-only inbox triage (in build, and we say so until it is proven)
Each workflow carries a deterministic gate that runs outside the model. A cited figure either traces to a source page or the output comes back flagged. The model does not grade its own work and neither do we. For the small set of tasks where a frontier model is still needed, a sanitization step strips client identifiers before anything leaves, and which tasks those are is a signed document the firm approves first.
The opinion's hardest requirement is the consent conversation. The simplest way to satisfy it is to never need it.
Questions firms ask
Can ChatGPT be used for confidential information at all? Under Opinion 512, only after you have assessed how the tool handles data and, for a self-learning tool, obtained the client's informed consent for that matter. For most firms, that makes the consumer product unusable for client work.
Does turning off training history fix it? It helps with Duty 1 and does nothing for the fact that the data still leaves your building and sits with a vendor under their terms. You still have to evaluate those terms and, depending on your bar, disclose.
Is an on-site model good enough for legal work? For the routine tier, reading, sorting, extracting, drafting first passes, yes, and it is where most of the hours are. For the hardest reasoning tasks a frontier model may still be better, which is why the escalation path exists and is sanitized and documented.