State Bar Rules on AI Use: What Lawyers Need to Know About AI Compliance

State bar rules on AI use are the rules of professional conduct a lawyer already answers to, applied to a new tool. Bar regulators across a growing number of jurisdictions, and the American Bar Association (ABA), have issued formal opinions or guidance saying so rather than writing a separate AI rulebook. California has gone further and become the first state to govern attorney AI use by statute.

This article covers five things. It names the governing instrument for each jurisdiction that has issued one. It covers California's new statute, the first in the country to regulate attorney AI use directly. It sets out the duties every opinion applies. It explains where AI use crosses into the unauthorized practice of law. And it covers how courts have enforced the standard so far against the lawyers and law firms that signed the filings.

Key takeaways

1. The state bar rules on using AI as a lawyer are the existing rules of professional conduct applied to generative AI. No bar regulator has written a separate AI rulebook, and most of what has been issued is advisory rather than disciplinary.
2. California is the exception. Senate Bill 574 adds Business and Professions Code section 6068.1 and makes California the first state to govern attorney AI use by statute instead of by guidance.
3. Every instrument issued so far puts the duty on the signing lawyer: verify what the tool produces, keep what goes into it out of reach of anyone outside the representation, and retain the legal judgment rather than delegating it.
4. Enforcement has come mostly from the bench rather than from bar counsel. Judges have imposed monetary sanctions, struck filings and referred lawyers to disciplinary authorities over briefs carrying invented citations.
5. A jurisdiction that has issued nothing is not a jurisdiction without rules. ABA Formal Opinion 512 and the lawyer's own rules of professional conduct still apply there.

[cta-1]

How state bar associations are approaching AI use guidelines for attorneys

State bar associations are approaching AI use guidelines for attorneys by applying the existing rules of professional conduct to a new tool rather than writing a separate AI rulebook. The table below names each jurisdiction that has issued guidance, its instrument and its date.

Why advisory guidance still decides outcomes

Only one instrument in the table carries disciplinary force. The instrument for the state a matter is filed in is what a regulator measures a lawyer's conduct against, whether or not it binds, so it is worth reading in full rather than in summary. The instrument names for Kentucky, Mississippi, Missouri, New Jersey, New Mexico, New York and West Virginia come from a state-by-state roster of bar AI guidance verified as of April 2025, which lists each instrument by name but carries no date for any of them; this page therefore states only the year for those rows rather than a more precise date no reachable authority supplies. Pennsylvania's and New York City's opinion documents carry no issuance date either, so their rows likewise show only the year each opinion's own number carries.

Where one jurisdiction needs two rows

California appears twice because its guidance and its statute are different instruments with different force. New York appears twice because the two instruments come from different bodies. The state-level output is the State Bar Association task force report. The New York City Bar opinion on AI recording and transcription comes from a voluntary city bar association that does not regulate admission to practice in the state.

What these instruments do not settle

Several gaps recur across these instruments:

1. They restate general duties more often than they prohibit a named workflow, so a lawyer looking for a rule about a specific tool usually finds a duty instead.
2. They address lawyers rather than vendors, because the rules of professional conduct bind admitted lawyers.
3. They predate agentic tools everywhere except California.

The District of Columbia opinion on generative AI in client matters is representative: competence, confidentiality and supervision in client representations.

What California state bar AI guidelines require, including agentic AI

California state bar AI guidelines now cover agentic AI, and California also became the first state to regulate attorney AI use by statute. The State Bar's Committee on Professional Responsibility and Conduct issued its 2026 Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law to replace the 2023 version, and it addresses agentic AI at the request of the California Supreme Court.

What the guidance adds for agentic AI

The guidance treats an agentic system as one that can pursue goals, plan, select tools and execute tasks without immediate human prompting. It then draws the line at substance: lawyers must not deploy such a system in a manner that allows it to make substantive legal determinations, communicate legal advice, prepare and file pleadings, or otherwise act in a representative capacity without meaningful lawyer supervision and review. It also treats configuration as an ethical act, because an agentic tool wired into firm email or client files holds persistent access a prompt-based tool does not, so the 2026 Practical Guidance asks a lawyer to limit the scope of that access.

What Senate Bill 574 requires

Senate Bill 574, signed September 30, 2026, adds Business and Professions Code section 6068.1 and amends Code of Civil Procedure section 128.7. Because the bill carries no urgency clause, it should take effect January 1, 2027. Section 6068.1 carries four duties:

1. No delegation of the practice of law to generative artificial intelligence.
2. No entry of confidential, personal identifying or other nonpublic information into a generative AI system unless access to those inputs is restricted.
3. Reasonable steps to verify the accuracy of AI output, including all case and statutory citations.
4. Disclosure of generative AI use to the court for all documents submitted to it.

A post-enactment analysis of Senate Bill 574 reads the verification duty as reaching any material the attorney uses, which pulls client memos and internal research inside a duty many firms have scoped to filings.

The court-disclosure duty is the operational change. Firms are likely to handle it with a standard disclosure paragraph, the way a word-count certification is handled, with a named owner on the matter team confirming whether AI was used before a document is filed. The court-facing side of that duty carries its own judge-specific rules.

Unauthorized practice of law: what AI tools state bar guidance says

State bar guidance on AI tools and the unauthorized practice of law draws one line: a lawyer may use an AI tool, and a lawyer may not delegate the practice of law to it. California wrote that line into statute, and Business and Professions Code section 6068.1 states that an attorney shall not delegate the practice of law to generative artificial intelligence. Florida reached the same conclusion through ethics guidance, holding that a lawyer cannot delegate to an AI tool the functions requiring a lawyer's personal judgment and participation.

Where delegation becomes the practice of law

Neither instrument defines delegation, which is where the practical difficulty sits. Florida Ethics Opinion 24-1 frames the test around personal judgment and participation rather than around which task was performed, so the question is not whether a tool produced something but whether a licensed lawyer made the legal calls in it. A first-pass clause the lawyer then revises is permitted. Advice sent to a client in substantially the form the tool produced is not.

Whether an AI provider is practicing law

The second half of the question is a question no state ethics opinion has yet answered: whether a provider selling legal output directly to consumers is itself practicing law without a license. That question runs through state unauthorized-practice statutes and the courts rather than through bar guidance. A commercial team should read that gap narrowly: the absence of a holding against a provider says nothing about the lawyer's own exposure.

[cta-2]

What does compliance with attorney ethics rules require for AI use?

Complying with attorney ethics rules when using AI requires a lawyer to satisfy duties that already existed, because no bar regulator has written a separate AI rulebook. The same duties carry the weight in every opinion issued so far:

1. Competence in the tool, which the opinions treat as understanding what the model does with an input and where it fabricates, not as technical expertise.
2. Confidentiality of what goes into it, tested by who can reach the input rather than by what the vendor advertises.
3. Verification of what comes out, including every citation and every factual summary.
4. Supervision of the people and the tools producing the work, reaching associates, paralegals and software alike.
5. Candor to the court about the result, including any disclosure the forum requires.
6. Reasonable fees for the time actually worked, which rules out billing the hours a tool saved.

A lawyer who can document those duties for a given tool is positioned to satisfy every jurisdiction that has issued guidance on it.

What a firm needs on record

Supervision is the duty a disciplinary inquiry is most likely to test on the record, so a firm needs to be able to show:

1. Who approved the tool, and what it was approved for.
2. Which lawyer reviewed the output on the matter in question.
3. What the tool was permitted to touch, including which systems it could reach.
4. Whether the output was changed before it left the firm.

Building that record is firm AI policy work rather than a per-matter decision.

Confidentiality: what state bar guidance says about client data and privilege

State bar guidance on generative AI, confidentiality and privilege applies one test: not whether a tool is encrypted, but whether anyone outside the representation can reach the input. North Carolina's 2024 Formal Ethics Opinion 1 disfavors putting client information into publicly available tools that retain and train on user input. California's Business and Professions Code section 6068.1 bars an attorney from entering confidential, personal identifying or other nonpublic information into a generative AI system unless access to those inputs is restricted to the attorney and to people the attorney has authorized who are themselves under obligations to protect the information.

What to settle with a vendor before approval

None of the opinions above requires an enterprise tier. They require that the input not be reachable outside the representation, which is a different test and one a procurement conversation can settle. Turning off model training is a start, and alone is often not enough, because a vendor that retains inputs for abuse monitoring or permits human review has put people outside the lawyer's control in reach of them. The questions that resolve that test sit in the contract rather than on the marketing page:

1. What is retained, and for how long.
2. Who at the vendor can read an input, and under what process.
3. Whether subprocessors are bound to the same terms.
4. Whether the terms survive termination of the subscription.
5. Whether the retention terms sit in the agreement rather than a policy page.

North Carolina's opinion on the use of artificial intelligence in law practice is the clearest statement of the public-tool problem.

Why privilege is the separate exposure

Disclosure to a vendor acting as the lawyer's agent in the representation is treated differently from disclosure to a third party, so a tool used outside that relationship can put privileged material with someone whose role no engagement document describes. The agreement, rather than the tool's feature set, is what a court would look at.

How firms require citation verification for AI-generated work

Law firms make citation verification a named step with an assigned owner rather than a general instruction to be careful. Every bar instrument on the subject places the duty on the signing lawyer: Texas Professional Ethics Committee Opinion 705 states that a lawyer should always verify the accuracy of any response received from a generative AI tool, and that AI output may serve as a starting point but must always be analyzed carefully for accuracy and quality.

What a verification record contains

1. The citation checked, against the source rather than against a search result.
2. The lawyer who checked it.
3. Whether the authority supports the proposition the document attributes to it.

Pennsylvania reaches the same place through a longer list. Joint Formal Opinion 2024-200 requires a lawyer who uses the technology to understand how it works, understand its benefits and risks, check and verify all citations and the material cited, and in some circumstances educate the client and seek informed consent. That duty reaches whether the authority says what the brief claims, which is the failure mode that survives a citation-checking tool and the one that reached a federal filing in the matter of the lawyer who used ChatGPT.

California now attaches a personal duty to the same step. The amended Code of Civil Procedure section 128.7 provides that a paper filed in any court shall not contain citations that the attorney responsible for submitting it has not personally verified, including any citation provided by generative artificial intelligence. Personally means that verification by a paralegal, a junior associate or a citation-checking tool does not discharge the duty.

What tasks can lawyers delegate to AI under current rules?

AI can draft legal documents, conduct research, review contracts and organize documents under current rules, provided the lawyer makes the legal judgments and verifies what comes back. No bar instrument lists which tasks are permitted; the opinions test the lawyer's involvement instead.

Tasks the opinions treat as ordinary with supervision

1. Drafting a legal document the lawyer then revises and stands behind, such as a first-pass clause or a routine amendment.
2. Legal research and case analysis, with every citation and holding verified before use.
3. Contract review, including first-pass risk identification against the firm's own positions.
4. Document and data organization across a large set, including diligence review.
5. Summarizing material the lawyer reads behind the summary.

Tasks that carry real exposure

1. Filing or serving anything without a documented human review of every citation.
2. Giving a client legal advice in substantially the form the tool produced.
3. Putting client confidential or personal identifying information into a consumer tool whose terms permit retention, review or training.
4. Letting an agentic tool take an external action, such as transmitting a document or contacting a court, without a lawyer approving it first.
5. Treating a tool's own citation check as the verification step the rules require.

The dividing line in both lists is whether a lawyer's judgment is in the output and can be shown to be. Common AI use cases in legal work cluster in the first list for that reason.

[cta-3]

AI legal malpractice developments since the ABA and state bar guidance

AI legal malpractice developments since the ABA and state bar opinions of 2024 have come, through 2025 and 2026, almost entirely from courts rather than from bar counsel. Judges have imposed monetary sanctions, struck filings and referred lawyers to state bar disciplinary counsel over briefs carrying citations that generative AI invented.

Why AI malpractice enforcement comes from the bench, not bar counsel

AI legal malpractice exposure since those opinions concentrates in court. The opinions set the standard a court measures conduct against, and so far the consequence has arrived as a court sanction within weeks of the filing rather than as a disciplinary proceeding. In 2026 the California Supreme Court referred a matter to State Bar counsel over briefing suspected of containing false AI-generated material, and a report on that referral sets out the sequence.

How advisory guidance shapes an AI legal malpractice claim

An advisory ethics opinion is not a disciplinary rule, but it is the document a plaintiff uses to establish what a reasonable lawyer should have done. How that plays out is visible in the sanctions record of the matter where a court fined lawyers over fabricated citations, because the inquiry moves from the brief to the lawyer's own conduct quickly.

Has a state bar issued guidance on ChatGPT for lawyers, and what applies if not

A lawyer checking whether a state bar has issued guidance on ChatGPT for lawyers often finds that it has not, and the answer changes nothing about the duties that apply. The rules of professional conduct apply in full in a state bar jurisdiction that has issued no AI guidance, and the national baseline for reading them is ABA Formal Opinion 512.

A state bar silent on ChatGPT has usually spoken on technology competence, cloud storage or outsourcing, and those opinions carry over.

Where to look when a state bar is silent

1. The state bar's ethics opinions index, including opinions on technology and outsourcing.
2. The state supreme court's rules of professional conduct and any AI order.
3. The local rules and standing orders of the court where the matter sits.

A lawyer whose state bar has issued no guidance on ChatGPT for lawyers should record that the check was made, because the absence of state bar guidance is unlikely to be treated as a defense.

What ABA Formal Opinion 512 supplies where a state bar is silent

Formal Opinion 512 was issued on July 29, 2024 by the Standing Committee on Ethics and Professional Responsibility, and it creates no new rules. It applies the existing Model Rules to generative AI use. The rules it reaches include competence, communication with the client, fees, confidentiality, meritorious claims, candor toward the tribunal, supervisory responsibilities and dishonesty. An analysis of Formal Opinion 512 as the paradigm for generative AI in legal practice is the practical way in, because the ABA document reads as a survey of rules rather than a checklist. A judge's standing order binds a filing whether or not the state bar has spoken, so it is the operative requirement in a silent jurisdiction.

How Spellbook helps lawyers meet bar ethics requirements

Spellbook is an AI platform for commercial legal work. The parts of it that bear on these duties are the controls and the record it keeps. Spellbook holds SOC 2 Type II, is compliant with the Health Insurance Portability and Accountability Act (HIPAA) and supports obligations under the General Data Protection Regulation (GDPR), and offers single sign-on, multi-factor authentication and encryption alongside zero data retention, meaning the model providers do not train on, learn from or store what a lawyer puts in. That is the test the confidentiality opinions apply: whether anyone outside the representation can reach the input. Work happens in Word and Google Docs, so review stays in the document a lawyer is already editing.

What the controls do

1. The Review feature surfaces risks and proposes redlines as native Word Track Changes, leaving a visible record of what the tool proposed and what the lawyer accepted.
2. The Draft feature drafts and auto-completes language using GPT-5, grounded in the firm's own positions rather than in open web text.
3. The Playbooks and Clause Library features hold those positions, so a first-pass review is measured against standards the firm has written down.
4. The Compare to Market feature checks document coverage against 2,300+ benchmark standards, with market positions drawn from 20M+ real contracts.
5. The Associate feature executes multi-step legal work across an entire document set from a single instruction, from due diligence to transaction document preparation, with the lawyer reviewing the output before it is used.

Where a platform stops

Spellbook is a commercial contract specialist rather than a legal research platform, so the citation-verification duty on a brief sits outside what it does, and no tool discharges a duty the rules place on the signing lawyer. A platform can lower the cost of the review and the record enough that the duty is met before a deadline, and help a team save up to 75% of time. Try Spellbook free to see the Review feature run against a contract.

Frequently asked questions

Are lawyers required to disclose AI use to clients?

In some circumstances, yes. Disclosure is generally required where AI use means sharing confidential information with a third-party vendor, or where it affects the basis of the fee, and the duty to keep a client reasonably informed about the means used supplies the general obligation. Several opinions also recommend obtaining informed consent before client information goes into a tool.

Which state has the strictest AI rules for lawyers?

California. It is the only state that regulates attorney AI use by statute rather than by advisory guidance, and the only one whose court-disclosure duty applies statewide rather than court by court.

Can AI practice law on its own under bar rules?

No. Under California's statute and the ethics guidance that has reached the question, a lawyer may not delegate the practice of law to a generative AI tool, and the functions requiring a lawyer's personal judgment and participation cannot be handed to software. A tool may produce material; the legal judgment and the responsibility for the result stay with the licensed lawyer.

Can a lawyer use a consumer chatbot for legal work under bar rules?

For work involving no client confidential information, usually yes, subject to the same verification duty as any tool. For anything carrying client information the problem is the terms: a consumer subscription whose terms permit the provider to retain, review or train on inputs puts people outside the representation in reach of them, which is a different proposition from the assistants now packaged for legal teams under enterprise terms.

Does a bar rule require disclosing AI use to a court?

In California, yes, by statute, for all documents submitted to the court. Elsewhere the requirement is forum-specific rather than statewide: judges have issued standing orders requiring disclosure, certification of verified citations, or both, so the operative source is the standing order in the court where the document is filed.

About this content: This article is provided for informational purposes only. It does not constitute legal advice, and reading it does not create an attorney-client relationship. For legal advice on a specific situation, consult a licensed attorney in your jurisdiction. Spellbook is a software product; it supports legal professionals' work but does not replace their judgment. Descriptions of ethics opinions, statutes and AI capabilities reflect publicly available information as of the publication date and may change.