AI Professional Ethics

State of play.

• Agentic AI has forced a governance model shift from reactive review to pre-deployment controls. Legal ethics commentary now frames the operative standard as "human-at-the-helm" — establishing parameters before autonomous action, not inspecting outputs after — with the EU AI Act and NIST AI Risk Management Framework increasingly cited as the regulatory backdrop .
• California is moving to convert advisory AI guidance into disciplinary-enforceable rules. COPRAC has proposed amendments to six Rules of Professional Conduct requiring independent verification of every AI output with no exceptions for routine matters — and has been directed to examine agentic AI implications as a next step .
• Client-side privilege waiver through consumer AI is now a documented judicial risk. United States v. Heppner (S.D.N.Y.) held that documents a client generated using a public AI platform are not privileged, and advisory guidance is now explicitly warning clients against uploading privileged materials to ChatGPT or Claude .
• ABA Formal Opinion 512 is the governing ethics baseline, but courts are imposing stricter standards through sanctions. The opinion's permission to reduce verification for "familiar" tools contradicts a Stanford study — cited in the opinion itself — finding legal AI hallucination rates of 17 to 33 percent .
• For counsel advising law firms or managing litigation teams, the practical baseline is that California's proposed binding verification rule signals where state bars are heading, ethics opinions set a permissive floor that courts are already exceeding through sanctions, and the client-side privilege risk now requires explicit engagement protocols before clients touch any public AI platform with sensitive materials.

Where things stand.

• ABA Formal Opinion 512 (July 2024) is the national ethics baseline. It requires technological competence under Model Rule 1.1, confidentiality protection, output verification, reasonable billing, and informed consent — and extends Rule 5.3 supervisory responsibility to AI-generated work product .
• State bars are adopting Opinion 512 verbatim, compounding its gaps. Mississippi Ethics Opinion No. 267 was adopted verbatim from Opinion 512, including the contested permission to reduce verification for familiar tools — a pattern likely to repeat in other jurisdictions .
• California is the leading jurisdiction converting advisory guidance into binding rules. COPRAC's proposed amendments to six Rules of Professional Conduct would impose enforceable verification obligations with no low-stakes exceptions, transforming the State Bar's November 2023 practical guidance into disciplinary-enforceable standards; the public comment period has closed and the proposal awaits final adoption .
• Fake-citation sanctions are a documented pattern across government and private practice. A Georgia prosecutor was suspended for AI-generated fake citations in a murder appeal; two New Orleans government attorneys resigned over the same issue; a Massachusetts attorney faced discipline; Flycatcher Corp. v. Affable Avenue produced a default judgment; and the 7th Circuit admonished a former immigration judge for citing fabricated cases .
• The privilege framework for AI-generated materials is unsettled at the district court level. SDNY's Heppner ruling turns on two factors — absence of attorney direction and public-platform confidentiality gaps — while E.D. Michigan's Warner v. Gilbarco found privilege intact where attorneys directed AI use without adversarial disclosure .
• Agentic AI governance is emerging as the next compliance frontier. The shift from generative to agentic systems — tools that send emails, populate filings, and modify records autonomously — renders post-hoc review inadequate; tiered risk management with pre-deployment controls is the framework now being advocated, with significant governance gaps remaining around data access sprawl and permission accumulation .
• Discovery workflows face a judicially imposed human-judgment floor. White v. Walmart (S.D. Ind., April 14, 2026) established that AI cannot satisfy the attorney's independent obligation to review, narrow disputes, and meet and confer in good faith .
• EDRM has published an embedded-safeguards framework as a competence benchmark. The guidance argues that training alone is insufficient — safeguards must be built into tools and must function under high-pressure conditions; Thomson Reuters is marketing "fiduciary-grade" AI as a response to this standard .
• Judicial AI adoption is broad. A Northwestern study found over 60 percent of surveyed federal judges report using AI in their work, raising questions about disclosure norms and the judiciary's own governance obligations .

Latest developments.

• Legal ethics commentary frames "human-at-the-helm" governance as the emerging standard for agentic AI — pre-deployment parameter-setting and tiered risk controls, not post-hoc output review — with significant governance gaps remaining around data access sprawl and permission accumulation .
• Advisory guidance from Varnum LLP and the National Law Review explicitly warns clients against uploading privileged documents to consumer AI platforms, citing Heppner and the absence of confidentiality obligations on public platforms; FTC injunctions against "robot lawyers" and state laws in Pennsylvania and New York restricting AI impersonation of licensed professionals are cited as the tightening regulatory backdrop .
• California COPRAC's proposed amendments to six Rules of Professional Conduct would impose binding verification obligations on every AI output with no routine-matter exceptions, converting the State Bar's 2023 advisory guidance into enforceable disciplinary standards; the rulemaking also directs examination of agentic AI implications .

Active questions and open splits.

• Does the ABA's "familiar tool" verification reduction survive scrutiny? Opinion 512's permission to reduce independent review for tools attorneys know well is internally inconsistent with the hallucination data the opinion itself cites — and California's proposed rule rejects the exception entirely, creating a direct conflict between the ABA baseline and the emerging state standard .
• Will the Heppner/Warner privilege split be resolved at the appellate level? The two district courts reached opposite conclusions on nearly simultaneous facts; the determinative variables — attorney direction, platform confidentiality terms — are not yet settled doctrine, and clients using public AI platforms remain exposed until a circuit weighs in .
• How far does the AI-driven discovery prohibition extend? White v. Walmart addressed meet-and-confer obligations, but the same logic could reach document review, privilege logging, and proportionality assessments; no court has drawn those lines yet .
• What governance architecture satisfies the duty of competence for agentic AI? The "human-at-the-helm" framework calls for pre-deployment controls and tiered risk management, but no bar or court has specified what those controls must look like — and California's rulemaking has only directed COPRAC to examine the question, not answered it .
• What does Rule 5.3 supervisory liability look like for AI in practice? Courts are holding supervising attorneys personally accountable for AI output, but the specific oversight protocols that satisfy the duty remain undefined — creating exposure for any firm that has not documented its review workflow .
• Are state bars moving faster or slower than courts? California's proposed binding rule moves ahead of most bars, while Mississippi adopted Opinion 512 verbatim without independent analysis — the gap between the most and least demanding state standards is widening, and courts are imposing stricter requirements through sanctions regardless of where bars land .

What to watch.

• Whether the California Supreme Court adopts COPRAC's proposed amendments and whether the final rule retains the no-exceptions verification standard — the first binding state ethics rule on AI will set the national benchmark.
• Whether any circuit court takes up the Heppner/Warner privilege split — the first appellate ruling will set the standard for platform selection and client counseling firm-wide.
• Whether additional state bars adopt Opinion 512 verbatim or begin diverging with stricter verification requirements in response to both the sanctions pattern and California's rulemaking.
• Whether bar disciplinary bodies or courts begin specifying what "human-at-the-helm" governance for agentic AI must look like in practice — the governance gap is currently self-defined by firms.
• Whether the White v. Walmart holding extends to other discovery contexts — privilege logging, document review, proportionality analysis — in follow-on decisions.
• Whether client-side privilege waiver incidents — driven by consumer AI use without counsel direction — begin generating malpractice claims against firms that failed to instruct clients on platform risks.