AI in Legal Practice: State-by-State Regulation and Bar Guidance (2026)

AI legal regulation by state is a patchwork, not a single rule. ABA Formal Opinion 512 (July 29, 2024) is the closest thing to a national baseline, and on top of it a handful of states have issued their own bar ethics opinions or court policies on lawyers using generative AI. Most states have not, and rely on the rules they already had. As of June 2026, no state bans generative AI for lawyers, and no state lets you skip checking what it produces.

If you run a legal team, the practical question is narrow. What has your state said, and what have the judges you appear before ordered? Those are two different layers, and the second one moves faster than the first.

One warning before you trust any AI-law tracker, including the AI-generated ones now ranking for this exact search. They invent bar opinions. While building this, we caught fake citations passed around as real: an "Iowa Ethics Opinion 24-01," a "Utah Advisory Opinion 24-03," an "Oregon 2024-197," none of which exist in any bar's records. AI is unusually good at fabricating this kind of thing because a plausible opinion number and date reads exactly like a real one, and few readers check. So every entry below is hand-verified against the bar or court's own source, and the fakes are left out.

The state-by-state table below is the artifact. Around it is the part that does not change much state to state: the duties every opinion repeats, and the court-level disclosure trend that is doing most of the live work.

TL;DR

• ABA Formal Opinion 512 (July 29, 2024) is the baseline. It applies the existing Rules of Professional Conduct to generative AI: competence, confidentiality, communication, candor, reasonable fees, and supervision. It is guidance, not law, but most state opinions track it closely.
• The count, by type. Nine jurisdictions have a numbered or formal state-bar ethics opinion on lawyer AI use (Florida, Pennsylvania, Texas, DC, Virginia, Kentucky, North Carolina, Oregon, Washington, with DC a district, not a state). The NYC Bar adds a city-bar opinion, and California has Board-approved practical guidance rather than a numbered opinion. Separately, four court systems have issued court-level guidance (New Jersey, Illinois, California's Judicial Council, Delaware). The rest is lighter still: a task-force report (NYSBA), a toolkit (Georgia), informal guidance (Minnesota), a non-binding guide (Ohio, new in 2026), and one judicial-conduct opinion (Michigan). Everything else is existing rules only.
• Nearly every opinion says the same five things. Understand the tool, protect client confidences, verify every output, charge a fair fee, and supervise the AI like a junior associate. None of them is a surprise if you already follow the rules.
• The court layer is where disclosure rules live. Individual judges started requiring AI-use certifications in May 2023, and well over 100 federal and state judges now have a standing order touching AI in filings. They are judge-by-judge, not uniform, and the Fifth Circuit declined to adopt a court-wide rule in June 2024.
• State court systems are starting to act. Illinois set an AI policy effective January 1, 2025, and California's Judicial Council adopted Rule of Court 10.430 effective September 1, 2025. Both let courts permit AI under existing rules rather than banning it.
• The sanctions risk is real, and it is the verification rule, not an AI rule. Fake AI-generated citations have drawn sanctions since Mata v. Avianca in 2023, and the cases have kept coming. Our sanctions tracker keeps the running count and the dollar figures, drawn from court decisions and the public AI-hallucination case databases that catalog them.

This is general information, not legal advice. Bar guidance changes fast, and several states updated their stance inside the last year, so treat this as a living tracker and confirm the current text before you rely on it.

The baseline every lawyer is under

Start with the document that covers all 50 states by default. ABA Formal Opinion 512, issued July 29, 2024 by the ABA Standing Committee on Ethics and Professional Responsibility, is the only national bar guidance on lawyers using generative AI. It does not create new rules. It reads the rules you already have onto a new tool. (The ABA hosts the opinion itself; the linked summary is a law-library writeup that loads for everyone.)

Six duties carry the opinion, and they show up again in almost every state opinion below:

1. Competence (Rule 1.1). Know what the tool can and cannot do before you use it on client work, and keep up as it changes.
2. Confidentiality (Rule 1.6). Check how a tool stores and trains on your inputs before you put client information into it.
3. Communication (Rule 1.4). Tell the client when your use of AI is material to the work or the fee.
4. Candor to the tribunal (Rule 3.3). Verify every citation and quote before it goes in a filing.
5. Reasonable fees (Rule 1.5). Do not bill a client for time you did not spend, and pass tool costs through honestly.
6. Supervision (Rules 5.1 and 5.3). You are responsible for what the AI produces the way you are responsible for a paralegal's work.

The ABA opinion is persuasive, not binding. For how that works, see are ABA formal opinions binding, and for a full walkthrough of the duties, the ABA Formal Opinion 512 guide. For the broader compliance picture (state AI statutes, the EU AI Act, the federal framework), see our AI legal regulation state and federal update.

The authority hierarchy (read this before the table)

These documents do not carry equal weight. Before you trust any single entry in the tracker, place it on this ladder, from least to most binding:

1. ABA formal opinion. Persuasive only. It binds no lawyer until a state adopts it or a court relies on it. It is the default framework everyone borrows from.
2. State-bar ethics opinion. Advisory in most states (guidance your bar will look to in a disciplinary matter, but not law). A few are stronger: Virginia's was adopted by the state Supreme Court, and North Carolina's was formally adopted by the State Bar Council, so those carry real authority in those states.
3. State supreme court or court-system rule. Binding on the courts and the people in them. Illinois's policy and California's Rule of Court 10.430 sit here.
4. Individual judge standing order. Binding in that one courtroom, full stop. It can require a certification your state bar does not, and it controls your filing regardless of what your bar opinion says.

Here is the opinionated part, because the ladder alone does not tell you how to act. Two advisory documents you should treat as operationally mandatory even though they are not technically binding: your own state bar's AI opinion, because it is exactly what a disciplinary panel will measure you against, and any standing order from a judge you are appearing before, which is binding in fact for that filing. Ignore neither.

What should not drive your firm-wide policy: another state's non-binding toolkit or informal column, and the ABA opinion read in isolation. They are useful reference, not a compliance baseline, and writing your policy to Georgia's toolkit when you practice in Texas is effort spent in the wrong place. Build to your own jurisdictions plus the strictest verification rule you find, and treat everything else as background reading.

Methodology

We tracked, per jurisdiction, four things: a state-bar ethics opinion specific to generative AI, a state court-system AI policy or rule, a bar or court AI task force or formal guidance, and notable judicial standing orders on AI disclosure. To count as "specific AI guidance," a document had to name generative or artificial intelligence and address how lawyers (or judges) may use it. We confirmed each against a primary source (the bar or court site) or a reputable secondary source (ABA, LawSites, a state bar journal, a firm client alert) we read.

We did not count generic technology-competence rules, CLE programs, or articles that merely summarize the ABA opinion. We also drew a clear line by type, because a "toolkit" or an informal column carries far less weight than a numbered formal opinion or an adopted court rule. The table tags every entry so you do not over-read a non-binding item.

On the fake-opinion problem flagged above: we treated every unverified opinion number as fabricated until a primary bar or court source confirmed it, and dropped the ones that failed. If your state is in the existing-rules group below and you later find a real opinion, it is new, not something we skipped.

We update this when a state issues new guidance. Dates and binding status are stated as of June 2026.

AI legal regulation by state: the 2026 tracker

Two honest groupings. A set of states has issued something specific to generative AI. Most states have not, and rely on their existing rules of professional conduct plus the ABA framework. "No specific guidance" means no state-specific AI document was identified as of June 2026, not that AI is unregulated there. The existing duties still apply everywhere.

The Type column tells you the weight. Binding items are an adopted court rule or a court-system policy. A "formal ethics opinion" is advisory in most states (persuasive, not law) unless a court adopts it. Task-force reports, toolkits, and informal columns are guidance only.

States with specific AI guidance

States relying on existing rules

No state-specific generative-AI ethics opinion was identified as of June 2026 for the states below. Several have a bar committee, a CLE program, or a journal article, but no numbered opinion or court policy aimed at lawyers using AI. The ABA 512 duties and each state's existing rules of professional conduct still apply in full.

Alabama, Alaska, Arizona, Arkansas, Colorado, Connecticut, Hawaii, Idaho, Indiana, Iowa, Kansas, Louisiana, Maine, Maryland, Massachusetts, Mississippi, Missouri, Montana, Nebraska, Nevada, New Hampshire, New Mexico, North Dakota, Oklahoma, Rhode Island, South Carolina, South Dakota, Tennessee, Utah, Vermont, West Virginia, Wisconsin, and Wyoming.

Note on Delaware: it has a court-system policy (in the table above), but it binds only judges and court staff, so practicing Delaware lawyers still operate under the existing rules plus ABA 512.

What changed in 2025 and 2026

The first wave (2023 into early 2024) was advisory bar opinions: Florida, DC, Kentucky, Pennsylvania, the NYC Bar, plus New Jersey's court guidelines. They all said the same thing, that existing rules apply. The newer wave is different in two ways.

New states issued guidance. Texas (Opinion 705, February 2025), Oregon (Formal Opinion 2025-205, February 2025), and Washington (Advisory Opinion 2025-05, 2025) added bar opinions. Virginia issued LEO 1901, approved by its Supreme Court in November 2025. Ohio's Board of Professional Conduct put out a non-binding ethics guide on June 2, 2026, and Georgia's State Bar published a toolkit, updated February 2026, that expressly is not an opinion.

Some guidance shifted from advisory to binding. This is the line that matters. Most of the early documents were advisory, and most still are. But the court layer has started issuing binding rules: Illinois's policy took effect January 1, 2025, and California's Judicial Council adopted Rule of Court 10.430, effective September 1, 2025, the first statewide court-system AI framework in the country. Of the four court systems that have acted, New Jersey's are guidelines for lawyers with no new duties and Delaware's bind only court staff and judges, while Illinois and California reach the courtroom itself. Virginia's opinion is unusual among bar opinions because the state Supreme Court adopted it, which gives it more force than a typical advisory opinion.

So the clean split today: bar ethics opinions are mostly advisory and tell you how to practice, while court-system rules and individual judge orders are binding and tell you what you may file. The advisory layer is broad and stable. The binding layer is newer, narrower, and moving.

The common throughline

Read a dozen of these opinions back to back and the wording barely changes. That is the useful part: you do not need 50 different policies, you need one that hits the points every opinion makes. Five duties recur almost word for word.

Competence. Every opinion ties AI to Rule 1.1. You have to understand the tool well enough to use it responsibly, including its biggest failure mode: it produces confident, fluent, wrong answers. Oregon's opinion cites research that general models hallucinate on courts' core rulings a large share of the time, which is why competence and verification travel together.

Confidentiality. The recurring warning is about inputs. Putting client information into a tool that trains on it, or stores it loosely, can waive privilege or breach Rule 1.6. Florida, Oregon, and the California guidance all push the same move: check the terms and the data handling first, and get client consent before using an open or consumer tool with confidential data.

Candor and verification. This is the one with teeth. Every opinion requires you to independently verify AI output before it reaches a court or a client. Texas Opinion 705, Pennsylvania 2024-200, North Carolina's 2024 FEO 1, and DC Opinion 388 say it almost identically: the duty of candor under Rule 3.3 is the lawyer's, and the tool does not carry it. For how to do it, see how to verify AI legal citations before filing.

Reasonable fees. Two halves. You cannot bill a client for hours the AI saved you, and you cannot bill for your own time learning a general tool. Virginia's LEO 1901 adds the flip side that surprised people: you are not required to discount a reasonable, value-based fee just because AI made you faster.

Supervision. Treat AI like a junior associate whose work you sign. You own the output under Rules 5.1 and 5.3. No opinion accepts "the AI did it" as a defense, and several extend the duty to vendor due diligence on the tool itself.

Court-level rules and standing orders

The bar opinions tell you how to practice. The judges tell you what to file, and that layer moves faster and varies more.

The standing-order trend started with one judge. On May 30, 2023, Judge Brantley Starr (N.D. Tex.) issued the first standing order requiring attorneys to certify either that no part of a filing was drafted by generative AI, or that a human checked any AI-drafted text against real sources. Well over 100 state and federal judges have since issued their own orders, per court-order trackers that catalog them. They are not uniform. Some demand certification, some demand disclosure, many courts have nothing.

Appellate courts have been more cautious. The Fifth Circuit floated a court-wide AI-certification rule in late 2023, took comment, and in June 2024 declined to adopt it. Its reasoning is the throughline of this whole topic: existing rules already require accurate, truthful filings, and "I used AI" excuses nothing. A special rule was deemed unnecessary.

State court systems are starting to set policy. Illinois put an AI policy in place effective January 1, 2025 that permits AI use under existing rules and does not require disclosure in pleadings. California's Judicial Council adopted Rule of Court 10.430, effective September 1, 2025, the first statewide court-system framework: courts that permit generative AI must adopt a written use policy by December 15, 2025 covering privacy, anti-bias, and human review, with disclosure when content is entirely AI-generated. Delaware's Supreme Court adopted an interim policy on October 21, 2024, but it governs judges and court staff, not practicing attorneys.

The enforcement risk is the verification rule, not an AI rule. Sanctions for fake AI citations trace back to Mata v. Avianca (S.D.N.Y., June 22, 2023), where the court fined two lawyers and their firm for filing ChatGPT-invented cases. The cases have not stopped. For the running list and the dollar figures, see our AI hallucination sanctions tracker and the deeper AI hallucinations and sanctions piece.

How a multistate team should operate

If you sit in-house or run a small firm across jurisdictions, the wrong move is 50 separate playbooks. The opinions overlap enough that three artifacts cover the field.

One AI-use policy, written to the duties. Anchor it to the six ABA 512 duties and the strictest verification language in any state you touch. A policy that requires human verification of every AI output, bans client confidences in open tools, and assigns a named owner satisfies California, Texas, Florida, and the rest at once. Update it on a fixed cadence, not ad hoc.

One vendor-review checklist, run before any tool is approved. The opinions that go past competence (North Carolina, Washington, the California guidance) all push diligence on the tool. Your checklist: does it train on your inputs, where is the data stored, is there a confidentiality and security commitment in the contract, and is there an audit or conformity record. Approve tools centrally so individual lawyers are not each negotiating terms.

One judge standing-order check, run per filing. This is the step teams skip. Before every filing, check the standing orders for that specific judge and court, because the disclosure or certification requirement lives there, not in your state bar opinion. Build it into the filing checklist the way you already check page limits and formatting.

Worked example: a filing in Illinois, before a federal judge with an AI standing order

Say your team drafts a summary-judgment brief with AI help for a case in the Northern District of Illinois. Illinois's state court policy permits AI and requires no disclosure in pleadings, so a lawyer who only checked the state layer would file with no certificate. That is the trap. You are in federal court, and the individual judge may have a standing order. The order outranks the state policy for this filing.

The right sequence: confirm the assigned judge's standing orders before drafting, not after. If the judge requires an AI-use certification, you draft to that format, run your own citation check against a real database (not the AI's own claim that the cite is real), and file the certificate with the brief. The state policy told you AI is allowed. The judge's order told you what to attach. Both matter, and the judge's order is the one that gets your brief stricken if you miss it.

Worked example: vendor review is not one decision

The opinions push diligence on the tool, and the answer differs by product because the data terms differ. A quick read of four common tools:

• ChatGPT Team or Enterprise. General-purpose, not legal-specific. The business tiers commit not to train on your inputs by default, which is the key confidentiality term, but it has no legal citation grounding, so verification load is highest. Confirm the data terms in your contract, not the marketing page.
• CoCounsel. Built for legal, sits on a vendor that publishes security and data-handling commitments. Vendor review focuses on the contract terms and whether outputs are grounded in real authority you can check.
• Harvey. Aimed at firms and in-house teams, sold through sales with enterprise data terms. Diligence is the same checklist (training, storage, security commitment, audit record), run against the negotiated agreement.
• Lexis+ AI. Tied to a licensed legal database, which changes the grounding question because outputs are linked to real sources, but you still verify and still confirm the confidentiality terms for your inputs.

The point is that "is this AI tool ethical to use" is not a yes or no. It is the same four-question checklist run per tool, and a legal-specific tool with real citation grounding lowers your verification risk but does not remove the duty.

Put the three artifacts in place and the patchwork stops being 50 problems. It becomes one policy, one checklist, and one per-filing habit that holds up as more states issue opinions.

Start free trialBook a demo

FAQ

Is it legal for lawyers to use AI in 2026? Yes. No US state bans lawyers from using generative AI. ABA Formal Opinion 512 and every state opinion issued so far permit it, subject to the existing duties of competence, confidentiality, candor, fees, and supervision. The constant condition is that you verify the output.

Which states have issued AI ethics opinions for lawyers? Confirmed as of June 2026: nine jurisdictions have a formal or numbered state-bar opinion (Florida, Pennsylvania, Texas, DC, Virginia, Kentucky, North Carolina, Oregon, and Washington, with DC a district, not a state). The NYC Bar issued a city-bar opinion, California has Board-approved practical guidance, and New York also has a state-bar task force report. Separately, four court systems issued court-level guidance (New Jersey, Illinois, California's Judicial Council, and Delaware), and Ohio published a non-binding guide in June 2026.

Is ABA Formal Opinion 512 binding? No. ABA formal opinions are persuasive, not binding. They become enforceable only when a state adopts them or a court relies on them. Most state opinions track the ABA 512 framework closely, so following it is the safe default, but your own state's rules control.

Do I have to tell the court I used AI? It depends on the judge and the court. Many courts require no disclosure, and Illinois and New Jersey expressly do not. But well over 100 individual judges have standing orders requiring an AI-use certification or disclosure. Always check the standing orders for your specific judge before filing.

What is the most common requirement across all the state opinions? Verification. Nearly every opinion requires you to independently check AI-generated citations, quotes, and analysis before using them. It flows from the duty of candor to the tribunal (Rule 3.3) and is the rule that draws sanctions when ignored.

Can I bill a client for time saved by AI? No. You cannot bill for hours you did not work, and you generally cannot bill for time spent learning a general AI tool. You can pass through tool costs with informed consent. Virginia's LEO 1901 adds that you are not required to discount a reasonable, value-based fee just because AI made you faster.

What happens if AI invents a case in my brief? You can be sanctioned. Courts have fined lawyers, ordered them to notify clients and judges, and referred them for discipline over fabricated AI citations, starting with Mata v. Avianca in 2023. The duty to verify is the lawyer's, not the tool's.

My state is not on the list. Does that mean AI is unregulated for lawyers here? No. It means no AI-specific ethics opinion was identified there as of June 2026. Your existing rules of professional conduct and the ABA 512 framework still apply in full. The duties travel with you even when a dedicated opinion has not been written yet.

Last updated: June 2026. We keep this current as states issue new opinions, so bookmark it and check the dates against the source before you rely on a specific opinion.

How to put this into practice

The pattern across every layer is the same. The rules you already follow govern AI, the one new habit is verifying output before it leaves your hands, and the disclosure question belongs to the judge, not just the state bar. Build your policy to the duties, not to a single state's opinion, and it holds up as the patchwork fills in.

We build Vaquill AI, a legal AI suite for in-house teams, with research, drafting, and matter management in one place and citation checking built into the workflow. For choosing tools against the duties above, see the best legal AI tools for in-house counsel.

Arshita Anand

Co-Founder & CEO · Attorney

Arshita leads product and strategy at Vaquill, building the legal AI suite that solo, small-firm, and in-house US lawyers use to run a matter end to end.