TL;DR: There is no single "AI ethics rule" for lawyers — there's the ABA's Formal Opinion 512 (July 2024) as a national baseline, plus a growing patchwork of state-specific opinions that agree on the fundamentals (competence, confidentiality, supervision, candor, billing) but differ on specifics like disclosure to clients and whether recording a client meeting with an AI notetaker needs consent. Below: what the ABA actually said, nine states with opinions we verified line by line, the themes that hold everywhere regardless of jurisdiction, and a policy checklist you can adapt today.
Updated: August 2026
Why this is confusing (and why that's not your fault)
If you've searched "can I use ChatGPT for legal work" and come away more confused than when you started, that's a reasonable reaction. No state ethics rule mentions AI by name — the Model Rules of Professional Conduct were written decades before generative AI existed. What's happened since 2023 is that bar associations have issued advisory opinions mapping existing rules (competence, confidentiality, supervision) onto AI tools, one jurisdiction at a time, on their own schedule. Some states have published detailed, numbered opinions. Others have task force reports, CLE guidance, or nothing formal yet. None of it is harmonized, and a rule that's settled in Texas may still be an open question in your state.
This isn't legal advice — always check your own jurisdiction's current opinions and your malpractice carrier's position before relying on anything here. What follows is a plain-language map of where things stood as of publication, verified against the underlying opinions rather than secondhand summaries.
The national baseline: ABA Formal Opinion 512
On July 29, 2024, the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, the first comprehensive national guidance on generative AI. It doesn't create new rules — it applies six existing Model Rules to GAI use:
| Rule | What it means for AI in practice |
|---|---|
| 1.1 — Competence | You don't need to become an engineer, but you do need to understand a tool's capabilities and limitations well enough to supervise its output — including its tendency to fabricate citations. |
| 1.6 — Confidentiality | Before pasting client facts into a tool, know whether that vendor trains its models on your inputs, where data is stored, and whether the terms of service actually protect privilege. "Free" consumer tools are the highest-risk category here. |
| 1.4 — Communication | Clients don't automatically need to be told you used AI to draft a first pass of a memo — but if AI use is material to the engagement (e.g., it affects strategy, cost, or an outcome), it may need to be disclosed. |
| 5.1 / 5.3 — Supervision | Partners and firm leadership need an actual written policy on permitted AI use, and need to supervise associates' and staff's use of it — "I didn't know they were using it" is not a defense. |
| 3.3 — Candor to the tribunal | Every citation, quote, and case cite generated by AI has to be independently verified before it goes in a filing. This is the rule behind every "lawyer sanctioned for fake AI citations" story. |
| 1.5 — Reasonable fees | If AI cuts the time a task takes, billing the client for the old, pre-AI hours isn't defensible. Value-based or flat fees sidestep this; hourly billing doesn't get a pass just because "that's how we've always billed it." |
Opinion 512 is advisory, not binding, but it's the framework nearly every state opinion since has either echoed or built on. If you read one source document, read this one.
State-by-state: what's actually been published
As of mid-2026, trackers maintained by the NYC Bar Association and other legal-ethics research groups count several dozen states with some form of published AI guidance — task force reports, CLE materials, or formal numbered opinions. Coverage and detail vary widely by state, and new opinions are issued every few months, so treat any fixed count as a snapshot, not a ceiling. Below are nine jurisdictions with opinions specific and citable enough to be genuinely useful, rather than a generic 50-row table padded with "no guidance yet."
| Jurisdiction | Opinion | Date | What makes it distinct |
|---|---|---|---|
| Florida | Ethics Opinion 24-1 | Jan 24, 2024 | One of the first out of the gate. Requires disclosure to clients when AI use affects billing or costs — a narrower, more specific disclosure trigger than most states adopted. |
| Kentucky | KBA E-457 | Mar 15, 2024 | Frames AI competence as an extension of existing technology-competence duties rather than a standalone obligation — useful if your state hasn't weighed in and you want a defensible analogy. |
| New York (NYSBA) | Task Force on AI report | Apr 2024 | Not a binding opinion, but the most detailed practical analysis published by any state task force — worth reading even if you don't practice in New York. |
| Pennsylvania / Philadelphia | Joint Formal Opinion 2024-200 | Jun 2024 | A rare joint opinion from a state and city bar together, which gives it broader practical weight across the Philadelphia market specifically. |
| North Carolina | 2024 Formal Ethics Opinion 1 | 2024 | Directly usable as a policy template — the opinion itself walks through what a firm AI-use policy should cover, not just what's prohibited. |
| New Jersey | Judiciary AI Guidelines | 2024 | Issued at the court-system level rather than as a bar ethics opinion — relevant if your matters involve NJ courts specifically, separate from your home-state bar rules. |
| Texas | Opinion 705 | Feb 2025 | Explicitly addresses the fabricated-citation problem and requires human verification of AI-generated legal analysis before it's relied on or filed. |
| New York (NYSBA) | Formal Opinion 2025-6 | 2025 | Narrow but practically important: covers AI notetakers/transcription tools in client meetings — consent and confidentiality obligations when a bot is "in the room." |
| California | COPRAC proposed Rule of Professional Conduct amendments | Approved for comment Mar 13, 2026 | Not yet final — this is a proposed rule change (following a 2023 non-binding practical guidance document), currently in a public comment period. Watch this one; it may become the first state to hard-code AI duties directly into its Rules of Professional Conduct rather than issuing an advisory opinion. |
If your state isn't listed above, that doesn't mean anything goes — it means the ABA baseline (and your state's existing competence/confidentiality/supervision rules) still apply in full; there just isn't state-specific commentary yet. Check your state bar's ethics opinion database directly, since new opinions are published regularly.
The five things every opinion agrees on
Strip away the jurisdiction-specific details and the same five points show up in nearly every opinion published so far:
- Verify everything AI generates before it touches a filing. Every sanctioned-lawyer story starts with skipping this step. No opinion treats "the AI said so" as a defense.
- Know your tool's data policy before you paste in client facts. Free consumer chatbots and enterprise legal-AI platforms are not the same risk category — check whether inputs are used for model training, and where.
- Someone senior has to own the policy. Supervisory duties under Rules 5.1/5.3 don't evaporate because the associate who used AI "should have known better." Firm leadership needs a written policy, not just a hallway conversation.
- Billing has to reflect actual time and value, not pre-AI habits. If a task that took three hours now takes forty minutes, hourly billing for three hours needs a real justification.
- Disclosure obligations are more likely to attach to outcomes than to tools. Most opinions don't require telling a client "I used ChatGPT to brainstorm an outline." They're much more focused on disclosure when AI use affects cost, strategy, or is material to the result.
A firm AI-use policy, in plain terms
North Carolina's opinion is unusually practical because it doubles as a policy outline. Adapted and expanded, a defensible one-page policy covers:
- Approved tools list. Name the specific platforms staff are permitted to use for client work, and explicitly ban pasting client-identifying facts into consumer tools not on the list.
- Verification requirement. Every citation, quote, and factual claim from AI output gets checked against a primary source before it's relied on — no exceptions for "obviously correct" results.
- Supervision assignment. Name who reviews AI-assisted work product before it goes out, and at what stage (draft review, not just final sign-off).
- Client disclosure trigger. Define the specific circumstances that require telling a client AI was used — tie it to cost/billing impact and material effect on strategy, matching the more concrete state standards (like Florida's).
- Logging. Keep a record of which matters used AI-assisted work and for what task — useful for both malpractice defense and for actually measuring whether the tools are paying for themselves.
Where to check for updates
This space moves in months, not years — a state that has "no guidance yet" today may have a formal opinion by year-end. For ongoing tracking beyond what's summarized here, the NYC Bar Association maintains a running compendium of opinions and reports across jurisdictions, and most state bar websites publish their own ethics opinion archives searchable by keyword. When in doubt, search "[your state] bar association artificial intelligence ethics opinion" directly rather than relying on any single third-party summary — this one included.
Bottom line
If your firm doesn't have a written AI policy yet, that's the actual risk — not which specific state opinion applies. Every jurisdiction that's weighed in agrees on the same core duties: verify output, protect confidentiality, supervise juniors, bill honestly, and disclose when it matters. Get those five things right in writing, and you're in a defensible position regardless of which state's opinion eventually lands closest to your fact pattern.