Lawyers who use generative AI in practice tend to use it in the same few places: getting oriented on an unfamiliar issue, producing a first draft, condensing long documents, and translating legal language for clients. The ethics rules don’t forbid any of that. They do shape how it has to be done, and the clearest statement of those limits at the national level is the American Bar Association’s Formal Opinion 512, issued July 29, 2024.
ABA opinions interpret the Model Rules. They aren’t binding on their own, and each state’s version of the rules governs. Some state bars have published their own guidance, so lawyers should read the opinion alongside whatever their jurisdiction has issued.
Where AI shows up in legal work
Orientation research
A lawyer facing an unfamiliar area often starts with a question to a chat tool: what issues does this raise, what kinds of authority usually apply, what facts matter? Used that way, AI works like a smart index. It points to places to look and doesn’t replace the looking.
I'm a lawyer getting oriented on [ISSUE] in [JURISDICTION].
List the legal issues this typically raises, the types of authority I should look for (statutes, regulations, leading case lines), and the facts that usually decide the outcome.
Do not cite specific cases. I will run the research in a citator-backed database.The last line is deliberate. General chat tools can produce citations to cases that don’t exist. NIST’s generative AI risk profile calls this “confabulation”: confidently stated but false content. The duty to check falls on the lawyer.
First drafts of routine documents
Engagement letters, demand letters, discovery requests, cover emails and internal memos all follow patterns. AI can produce a structured first draft from a short brief, which the lawyer then rewrites.
Draft a first version of a [DOCUMENT TYPE] for a [MATTER TYPE] matter in [JURISDICTION].
Key facts (no client names): [FACTS]
Points that must be covered: [LIST]
Tone: [e.g. firm but not aggressive]
Mark every statement of law with [VERIFY] so I can check it.Summarizing large document sets
Deposition transcripts, document productions and contract portfolios are where legal-specific platforms focus. Firms usually run this work inside a vendor platform under contract rather than a consumer chatbot, because the inputs are client confidential by definition. Firms comparing those platforms can start with the best AI for lawyers.
Plain-language client communication
Explaining a ruling, a settlement offer or a process to a client is a writing task that AI handles well when given the substance.
Rewrite this explanation for a client with no legal background, in under 200 words:
[LAWYER'S OWN EXPLANATION, WITH IDENTIFYING DETAILS REMOVED]
Keep every qualification and deadline. Do not add anything that isn't in my text.The output is a draft. The lawyer sending it is responsible for every word.
What Opinion 512 requires
The opinion works through existing Model Rules rather than creating new ones. A summary by rule:
| Model Rule | What the opinion says about generative AI |
|---|---|
| 1.1 Competence | Lawyers need a reasonable understanding of the tool’s capabilities and limits, and relying on output without appropriate independent review can breach the duty |
| 1.6 Confidentiality | Before putting client information into a self-learning tool, the lawyer needs the client’s informed consent, and general boilerplate in an engagement letter isn’t enough |
| 1.4 Communication | Disclosure is needed when the client asks, when engagement terms call for it, or when AI output will influence a significant decision in the matter |
| 3.1, 3.3 Meritorious claims and candor | Filings must be reviewed so they don’t contain misstatements of law or fact |
| 5.1, 5.3 Supervision | Managers should set firm policies, train lawyers and staff, and vet vendors much as they would for cloud services |
| 1.5 Fees | Hourly billing covers time actually spent; lawyers generally can’t bill for learning a tool; tools built into ordinary software are overhead |
Two points in that table catch lawyers out most often.
Consent has to be informed. Opinion 512 ties informed consent to an explanation of the risk that later users of a self-learning tool could get access to information about the representation. A clause buried in the engagement letter doesn’t meet that standard. Tools that don’t learn from inputs, under terms the lawyer has read, are a different situation, which is why the opinion expects lawyers to understand the terms of use.
Faster work changes the fee picture. If AI makes a task much quicker, charging the old flat fee for it may be unreasonable under Rule 1.5. A pay-per-use AI service can be passed on at actual cost. An AI feature built into a word processor can’t be billed to a matter at all.
Where lawyers shouldn’t use it
Don’t paste this in
Client confidential information into any tool that learns from inputs, unless the client has given informed consent after a real explanation of the risk. ABA Formal Opinion 512 (July 29, 2024) says boilerplate engagement-letter language does not supply that consent. Privileged communications, sealed material, protective-order documents and opposing parties’ confidential productions carry their own limits on top of Rule 1.6.
Beyond confidentiality, a few uses are poor fits even with consent:
- Unchecked citations. Every case, statute and quotation needs to be pulled and read before it goes in a filing.
- Final judgment calls. The opinion says lawyers can’t hand off professional judgment to the tool. Settlement advice, case strategy and conflict decisions stay with the lawyer.
- Filings without a rules check. Courts and individual judges can set their own requirements, so checking local rules and standing orders for anything on AI use or disclosure belongs in the filing routine.
- Billing narratives for time not spent. Time entries must reflect real work.
A worked example: one matter, three tasks
Consider a small employment practice handling a wage claim for an individual client. Here is how the opinion’s logic plays out across one matter.
Task one: getting oriented. An associate asks a general chat tool for the issues a misclassification claim raises in the state. No client facts go in, so confidentiality isn’t engaged. The output is a list of questions and authority types, and the associate then runs the actual research in a legal database and reads each authority.
Task two: reviewing the client’s pay records. This is client confidential information. The firm runs it through a legal platform whose contract says inputs aren’t used for training and are deleted on a set schedule. Under Opinion 512’s reasoning, that’s a different risk from a self-learning tool. The lawyer has read the terms and noted them in the file, and the client was told how the review would be done.
Task three: the demand letter. A first draft comes from the platform, built from the lawyer’s own outline. The lawyer rewrites it, checks every statutory reference and decides the demand figure without the tool. The matter is billed hourly, so the time entry reflects the time actually spent reviewing and rewriting, not the time a draft would have taken by hand.
None of these steps is unusual. The difference from careless use is that each one has a decision attached: what data goes where, under which terms, and who checks the result.
A short firm checklist
- List approved tools, the account tier and what data each can receive.
- Read each tool’s terms on training, retention and access, and keep a dated note.
- Write an informed-consent explanation clients can actually follow, for matters where consent is needed.
- Require citation checks on anything filed or sent to opposing counsel.
- Update billing guidance for AI-assisted tasks and AI tool costs.
- Train staff, including paralegals and assistants, under Rule 5.3.
Lawyers who prepare tax returns have one more layer. Return preparers are subject to Section 7216 on tax return information, which is covered in the guide to AI for accountants.
Sources
- American Bar Association, Formal Opinion 512: Generative Artificial Intelligence Tools, July 29, 2024.
- Frankfurt Kurnit Klein & Selz, summary of Formal Opinion 512, used to cross-check the fee and supervision points.
- NIST, AI Risk Management Framework: Generative AI Profile, July 26, 2024, on confabulated output.