What the ABA and State Bars Require of Lawyers Using AI

The practical problem with artificial intelligence in the legal sector is no longer whether lawyers are experimenting with it. They are. The harder problem is that individual use has outrun firm-level controls: one 2026 legal industry survey reported that 69–79% of legal professionals use AI in some capacity, while only 9% of firms have an enforced written AI policy; the same report found that 44% of law firms have no formal AI governance policy at all, and 54% provide no training and have no plans to do so.[1]

That gap matters because professional responsibility rules do not become optional when a task moves from a junior associate, paralegal, search platform, or contract attorney into a generative AI interface. ABA Formal Opinion 512, issued in 2024, applies familiar duties—competence, communication, confidentiality, candor, supervision, reasonable fees, and the prohibition on unauthorized practice—to lawyers’ use of generative AI tools.[2] The rules are old. The failure points are multiplying.

This article is an educational reference synthesis, not legal advice. It does not substitute for checking the rules, ethics opinions, standing orders, local rules, client requirements, and vendor contracts that apply to a particular matter, court, jurisdiction, or organization. The source landscape is also moving quickly: state guidance, court disclosure orders, and reported AI-related sanctions should be checked against primary sources before a filing, engagement approval, or firmwide rollout.

ABA Formal Opinion 512 Is the Starting Point, Not the Whole Map

ABA Formal Opinion 512 is useful because it refuses to treat AI as a special ethics-free category. It does not ask whether a lawyer is excited or anxious about the technology. It asks whether the lawyer can competently use the tool, protect client information, communicate when necessary, supervise people and systems involved in the work, charge reasonably, and avoid misleading a tribunal.[2]

Its most operationally important move is the vendor due diligence expectation. A lawyer cannot safely assume that a product’s polished interface, legal-market branding, or generic terms of service resolve the confidentiality, supervision, or accuracy issues. Before relying on a tool, the lawyer needs to understand what the tool does with inputs, whether information is used for training, who can access it, what contractual confidentiality commitments exist, what security controls apply, and whether the output can be verified.

State bars have started filling in that framework. As of mid-2026, formal or substantial AI guidance exists in at least 12 jurisdictions or bar sources, including California, Florida, New York, Texas, Pennsylvania, Oregon, Kentucky, and New Mexico, with additional state-by-state tracking maintained by legal practice resources.[2][5] The guidance is not identical, but it tends to return to the same evidence question: if something goes wrong, can the lawyer show what was reviewed, who approved it, what was disclosed, and what was verified?

Competence Means Knowing the Tool, Not Just Knowing the Law

Model Rule 1.1 requires competent representation. In AI work, competence has two layers. The first is the ordinary legal layer: the lawyer still must know the doctrine, procedure, facts, and forum. The second is tool competence: the lawyer must understand enough about the AI system’s capabilities and limits to decide whether it is appropriate for the task and how much verification is required.

ABA Formal Opinion 512 treats that tool understanding as part of the lawyer’s duty, not as an optional technical preference.[2] A lawyer using generative AI for legal research, factual summaries, brief drafting, privilege review, deposition preparation, contract analysis, or client communications should know whether the tool is retrieving from a controlled database, generating probabilistic text, searching uploaded materials, or combining those functions. The review burden changes depending on the answer.

Sources cited here report that, as of Q3 2025, New York mandates at least two annual CLE credits in AI competency.[2] That requirement should not be read as a New York-only curiosity. It signals how quickly “I did not understand the tool” can become a weak answer when the tool is already part of legal work.

For a solo lawyer or small firm, competence documentation does not need to look like a global technology committee memo. It can be a short approved-use chart, a record of completed training, a list of prohibited inputs, a verification checklist for research and drafting, and a date-stamped vendor review. What matters is that competence is visible before the error, not reconstructed after a court asks what happened.

Confidentiality Is the Highest-Risk Duty

Model Rule 1.6 requires lawyers to protect information relating to the representation of a client. In AI use, that duty starts before the prompt is typed. The confidentiality question is not limited to whether a lawyer intended to disclose client information. It includes whether client information is transmitted to a third party, retained, reviewed, reused, used for model training, exposed through security weaknesses, or placed outside the protections needed for privilege and confidentiality.

ABA Formal Opinion 512’s vendor due diligence expectation belongs here first. Terms of service are not the same as a legal ethics review. A lawyer or firm should know whether the vendor contract contains confidentiality obligations, whether the vendor disclaims professional duties, whether prompts and uploads are stored, whether human review is possible, whether data is segregated, whether enterprise settings differ from consumer settings, and whether the lawyer can prevent client data from being used to train models.

A February 2026 Southern District of New York privilege ruling described in a secondary source, Heppner, is especially important if confirmed by the docket. The report states that documents created using generative AI without contractual confidentiality guarantees were held not protected by attorney-client privilege.[4] That is not a point to cite casually. Before treating Heppner as a privilege landmark, counsel should verify the court record directly. But the reported lesson is consistent with the direction of the ethics guidance: confidentiality cannot rest on hope, branding, or default settings.

The safer posture is contractual and documented. For client-confidential uses, the lawyer should be able to produce the vendor review, the operative contract or order form, the confidentiality provision, the data-use restriction, the security review, and the internal approval showing which matters and data categories may be used. If the tool cannot provide those assurances, the use case may need to be limited to non-confidential material, anonymized hypotheticals, or workflows that do not transmit protected information.

Client Consent Is Not a Cure-All

Client consent may be required for some AI uses, especially where confidential information is shared with a third-party provider or the use materially affects the representation. But consent does not replace competence or supervision. A client cannot meaningfully authorize a risk the lawyer has not investigated, and a consent email does not fix an output that no one verified.

Supervision Now Includes Systems, Vendors, and Policy Enforcement

Model Rules 5.1 and 5.3 require supervisory lawyers to make reasonable efforts to ensure that other lawyers and nonlawyer assistants comply with professional obligations. AI complicates that duty because the work may pass through a vendor platform, an internal automation, a staff member using an unapproved tool, or a lawyer who treats AI output as a draft from nowhere.

ABA Formal Opinion 512 places AI use within those supervision duties.[2] That means a firm’s obligation is not satisfied by telling lawyers to “be careful.” A workable policy should say which tools are approved, which uses are prohibited, what information may be entered, who must review outputs, when client consent is required, how court-specific disclosure checks happen, and what training is mandatory before use.

The sanctions record makes the governance issue concrete. In Mata v. Avianca, lawyers were sanctioned $5,000 after submitting filings containing hallucinated case citations generated through ChatGPT.[2] In Lnu v. Blanche, sources cited here identify a candor violation tied to AI-generated filings.[2] In Withers v. City of Aberdeen, the sanction was reportedly enhanced because the firm had no AI policy, making lack of governance an aggravating factor rather than a background embarrassment.[2]

Those cases should not be used as theater. They show what courts are likely to ask after a defective filing: who used the tool, who checked the authorities, what policy governed the use, what training existed, and why the supervising lawyer believed the final work product was reliable. If the only answer is that the lawyer trusted the output, the supervision problem is already visible.

Candor Requires Verification Before the Filing Leaves the Building

Model Rule 3.3 prohibits false statements to a tribunal and requires correction of certain false statements. AI does not change that rule, but it increases the odds that a false citation, nonexistent quotation, distorted holding, or unsupported factual assertion can look polished enough to pass a tired reviewer.

The verification duty should be task-specific. For legal research, every cited authority should be checked in a reliable legal research database or official source. For quotations, the quoted language should be matched to the source. For factual summaries, the output should be traced to the record. For procedural statements, the rule and local practice should be checked as of the filing date. For declarations or client-facing factual narratives, the person with knowledge should review the substance.

Secondary sources cited here report at least 486 documented AI hallucination cases before courts worldwide, including 324 in the United States, maintained by HEC Paris.[2][4] Those numbers should be verified against the primary database before reliance. Even treated cautiously, they support a narrow and practical conclusion: hallucinated legal material is no longer a novelty, and courts have enough examples to expect lawyers to know the risk.

Court Disclosure Rules Belong in the Filing Workflow

Court-specific AI rules should not live in a memo that no one sees after onboarding. They belong in the filing workflow, next to Rule 11 review, local-rule checks, judge-specific procedures, and client-service requirements. A lawyer preparing to file needs to know whether the court requires disclosure, certification, special verification, or no separate AI statement.

Secondary sources cited here refer to more than 681 court orders and rules on AI use tracked by Ropes & Gray.[4] That figure should be checked against the primary tracker before reliance, but the operational point is already clear: disclosure obligations vary too much to be handled by memory. A national practice, even a small one, needs a court-order check at the matter and filing level.

Examples cited here include Judge Michael Baylson’s standing order in the Eastern District of Pennsylvania and Florida Rule 2.515(d)(2), effective June 15, 2026.[2][4] The Florida rule is especially important because it moves the AI issue into a procedural certification environment. Lawyers filing in Florida courts should not assume that an internal verification process is enough if the rule also requires a particular certification or representation.

Federal Rule of Civil Procedure 11 already requires an attorney to certify that filings are legally and factually supportable after reasonable inquiry. AI assistance does not dilute that signature obligation. If anything, it makes the “reasonable inquiry” record more important: what was generated, what was checked, what sources were used, and who performed the final review.

Fees and Client Communication Are Quieter Issues, But Not Optional

Model Rule 1.5 requires reasonable fees. AI creates a billing problem when it materially reduces time spent but the billing model continues as if the old labor pattern still exists. The issue is not that lawyers must pass every efficiency gain through to the client in the same way. The issue is whether the fee remains reasonable and whether the billing description is accurate.

The 2026 legal industry survey reported that 86% of solo firms and 78% of small firms had not adjusted pricing models for AI efficiency.[1] That statistic does not prove overbilling. It does show that many firms may be adopting AI before deciding how AI-assisted work should be priced, described, or explained to clients.

Model Rule 1.4 adds the communication layer. Some AI use may be routine enough not to require a separate client conversation. Other uses may require disclosure or consent because they affect confidentiality, cost, staffing, strategy, or the means by which the representation is carried out. State guidance differs on where that line falls, so a firm policy should not treat communication as a single universal answer.

A useful engagement-letter approach is not to bury a broad AI waiver in boilerplate. It is to identify categories of permitted technology use, reserve client-specific restrictions, disclose third-party processing where needed, and create a process for heightened consent when confidential information, sensitive data, or material strategic judgment is involved.

State Guidance Is Converging, But the Details Still Matter

The state materials cited here do not create one national AI rule. California’s expanded May 2026 guidance, Florida Opinion 24-1 and Rule 2.515(d)(2), NYSBA task force recommendations, Texas Opinion 705, Pennsylvania Joint Opinion 2024-200, Oregon Formal Opinion 2025-205, Kentucky E-457, and New Mexico 2024-004 all sit within the same professional-responsibility frame, but they may differ in emphasis, disclosure expectations, and procedural consequences.[2][5]

For firms operating across jurisdictions, the safest policy design is modular. The core policy can set minimum firmwide requirements: approved tools, prohibited inputs, vendor review, training, supervision, verification, billing review, and documentation. Jurisdictional addenda can then handle state bar opinions, court-specific disclosure rules, client-sector requirements, and judge-specific standing orders.

That structure matters for in-house departments too. A legal department approving AI tools for contract review, litigation management, employment advice, or regulatory work should not rely only on enterprise procurement approval. Procurement may evaluate cost, security, and integration. Legal ethics review asks different questions: whether the tool use affects privilege, client confidentiality, unauthorized practice, supervision of outside counsel, filing obligations, or billing review.

Vendor Surveys Are Useful, But They Are Not Ethics Authority

Vendor-published and vendor-adjacent surveys are helpful for understanding adoption, but they should be handled with care. The 69–79% adoption range, the 9% enforced-policy figure, and the training-policy findings are useful because they describe a governance gap, not because they establish a professional standard.[1] Survey methodology, respondent pool, firm size, practice area, and the definition of “AI use” can shift the numbers.

The same caution applies to ROI claims. Thomson Reuters has reported that 81% of organizations with a visible AI strategy see ROI, compared with 23% without one.[3] That may support the business case for strategy, but it does not answer whether a particular legal use satisfies Model Rule 1.6, a court disclosure order, or a client’s outside-counsel guidelines.

Adoption is evidence of practice reality. It is not evidence of ethical sufficiency. A firm can be technologically current and still be careless if it has no vendor review, no confidentiality guarantees, no training, no supervision protocol, and no filing disclosure check.

What a Responsible AI File Should Contain

The difference between responsible use and wishful adoption is usually documentary. If a lawyer or firm is going to rely on AI in legal work, the record should make the risk analysis visible before anyone has to defend it.

• Vendor due diligence: the reviewed tool, version or service tier, data-use terms, retention practices, training-use settings, security review, and contract owner.
• Confidentiality protections: contractual confidentiality language, restrictions on model training, access controls, data-location or retention terms where relevant, and client-consent analysis.
• Training records: who may use the tool, what training they completed, what use cases are approved, and what uses are prohibited.
• Supervision protocol: who reviews AI-assisted work, what must be independently verified, when escalation is required, and who owns final approval.
• Filing checklist: citation verification, record verification, Rule 11 review, local rule review, judge-specific AI order check, and required disclosure or certification.
• Jurisdictional source log: state bar opinions, court orders, procedural rules, last-reviewed date, and person responsible for updates.

A large firm may implement that record through a risk committee, knowledge-management system, and centralized procurement workflow. A solo lawyer may implement it through a controlled checklist and a folder of reviewed source materials. The scale can differ. The duties do not disappear.

The Maintained-Reference Posture

AI compliance for lawyers is not a one-time policy project. State bars are issuing new guidance, courts are adopting and revising disclosure orders, vendors are changing terms, and reported sanctions are giving judges a more concrete vocabulary for failed supervision. A policy approved in January can be stale by July if no one is assigned to maintain it.

The final professional-responsibility point is simple, but not soft: a lawyer cannot outsource ethics to a product interface, a vendor FAQ, or a chatbot’s default terms. Compliance requires evidence—vendor diligence, confidentiality guarantees, training records, supervision protocols, disclosure checks, verified filings, and current jurisdictional sources. That is the record a lawyer wants in place before the court, client, insurer, regulator, or managing partner asks what actually changed when the firm started using AI.

References

1. Legal Industry Report 2026, 8am.
2. AI for Lawyers: Ethics and Legal AI, Clio.
3. How AI is transforming the legal profession, Thomson Reuters.
4. Legal AI Trends 2026, Legartis.
5. By the Numbers: What Surveys Show About Law Firm AI Adoption, North Carolina Bar Association, May 2026.