AI Agents for Lawyers

Regulation & Ethics

ABA Formal Opinion 512 Explained: What US Lawyers Must Do When Using AI

ABA Formal Opinion 512 applies six Model Rules to generative AI: competence, confidentiality, communication, candour, supervision and fees. What each duty requires in 2026.

By Dr. Niklas SchmidtUpdated 13 November 20268 min read
On this page
  1. The six duties in one table
  2. Competence: what "reasonable understanding" means in 2026
  3. Confidentiality and the informed-consent question
  4. Fees: time, subscriptions and the "time saved" trap
  5. How 2026 guidance builds on 512
  6. Mapping Opinion 512 to a firm policy
  7. What to do this month

ABA Formal Opinion 512, issued on 29 July 2024, tells US lawyers that six existing Model Rules govern their use of generative AI: competence (1.1), confidentiality (1.6), communication (1.4), candour to the tribunal (3.3), supervision (5.1 to 5.3) and reasonable fees (1.5). It created no new rule. As of June 2026, seventeen jurisdictions plus the ABA have formal AI opinions, and all of them build on this frame.

The six duties in one table

The opinion itself is short, and its operative content fits in a table.

Model RuleWhat Opinion 512 requiresWhat it looks like in practice
1.1 CompetenceA reasonable understanding of the tool's capabilities and limits, kept current; independent verification of outputKnow that the tool invents citations; read every authority before filing; re-read the vendor's data terms when the plan changes
1.6 ConfidentialityEvaluate the risk that information is disclosed to or used by others; obtain informed consent before feeding client information into a self-learning tool; boilerplate consent is not enoughConsumer accounts with training on are off-limits for client data; enterprise contracts with no-training terms are evaluated and documented
1.4 CommunicationTell the client about AI use when it is relevant to the representation, when the client asks, or when consent is requiredEngagement letter paragraph plus a conversation where the tool touches strategy or confidential facts
3.3 Candour; 3.1 Meritorious claimsDo not submit fabricated authority; correct the record if you doA citation-verification step before signature, without exception
5.1–5.3 SupervisionFirm policies and training for lawyers and non-lawyer staff; AI treated like a non-lawyer assistant whose work must be checkedWritten policy, onboarding module, partner review of AI-assisted drafts
1.5 FeesBill actual time, not time saved; no charge for learning the tool unless the client requested it; subscriptions are overhead unless agreed; per-use costs passed through with disclosureTime entries describe the work done, not "AI"; engagement letter addresses AI costs

The ABA's own press release summarised the opinion as requiring lawyers to "fully consider their applicable ethical obligations". The Business Law Today summary adds the detail most readers miss: use of the same tool by several lawyers in a firm can itself lead to inadvertent disclosure between matters, which is why enterprise deployments need matter-level separation.

Competence: what "reasonable understanding" means in 2026

Opinion 512 does not require lawyers to become technologists. It requires enough understanding to evaluate the risk of the specific tool for the specific task, and it requires that understanding to be kept current. In 2024 that meant knowing chatbots fabricate cases. In 2026 the tools have memory, run agents that take actions, connect to document management systems and browse the web. The Texas Committee on Professional Ethics put the standard in one sentence in Opinion 705 (February 2025): a lawyer "must have a reasonable and current understanding of the technology", and "cannot blindly rely upon" its answers.

Two things make competence concrete. First, verification. The State Bar of California's May 2026 guidance states the rule for filings without qualification: "No paper filed in court should contain citations that the attorney responsible for the filing has not personally read and verified", as Wilson Elser summarised it. A citation-checking protocol is the operational form of Rule 1.1. Second, currency. A firm whose AI training dates from 2024 and does not cover agents, memory or connectors cannot show its lawyers understand the tools they now use; the hallucination guide is the minimum syllabus.

Rule 1.6 is where Opinion 512 goes furthest. Its logic: a self-learning tool stores inputs and may reproduce them for other users; putting client information into such a tool risks revealing it; revealing client information requires informed consent; and consent buried in an engagement letter is not informed. The proposed California comment to Rule 1.6, approved by COPRAC on 13 March 2026, pushes the same idea into rule text by treating exposure of client data to an AI system as "revealing" it where there is "a material risk of unauthorized access, retention, or use".

The practical division is between tools, not tasks.

  • Consumer accounts are self-learning in the ABA's sense as of September 2026: ChatGPT Free, Plus and Pro train on conversations unless the user opts out; Claude Free, Pro and Max require the user to choose, and opted-in data is kept for five years; consumer Gemini keeps activity for eighteen months by default and uses it for training unless the setting is off. Client confidences require consent, and even with consent they are a poor choice. United States v. Heppner (S.D.N.Y., 17 February 2026) showed the further problem: Judge Rakoff held that memos a defendant produced with consumer Claude were not privileged, because "Claude is not an attorney" and the vendor's terms allow logging and disclosure, as the Harvard Law Review Blog reported. The privilege page works through the consequences.
  • Enterprise and API deployments (Claude Team and Enterprise, ChatGPT Business and Enterprise, Harvey, Legora, CoCounsel) contractually exclude training. The duty is competent evaluation of the terms, retention periods and security, documented once, rather than matter-by-matter consent. The confidentiality checklist lists the terms as of September 2026, and the zero data retention entry explains the strongest setting available.

Opinion 512 also expects disclosure to the client under Rule 1.4 in some circumstances even where consent is not required. The client-disclosure page compares the state positions, which diverge more here than anywhere else.

Fees: time, subscriptions and the "time saved" trap

Opinion 512's billing rules are the ones clients now ask about in RFPs.

ChargePermitted under 512Condition
Hourly time spent prompting, reviewing and correcting AI outputYesMust be time actually spent; entries should describe the legal work
Hourly time the tool savedNoA task that took twenty minutes bills twenty minutes
Time spent learning a toolNoUnless the client asked for that specific tool
General AI subscription (Claude Team, Copilot, Harvey seat)OverheadRecoverable only if disclosed and agreed in advance
Per-use or per-matter AI chargesPass-through at costDisclosed in advance; no mark-up
Flat or value fee that reflects efficiencyYesReasonable under Rule 1.5(a); Virginia LEO 1901 (approved 24 November 2025) confirms a value-based fee need not be cut because AI saved time

California's 2026 guidance adds a hard edge: general AI subscriptions are overhead, and there is no mark-up on AI costs without informed written consent. Texas Opinion 705 states that a lawyer "may not charge hourly fees for the time that was 'saved'". The billing page has engagement-letter language for each model.

How 2026 guidance builds on 512

Opinion 512 was written before agents. The state bodies that have moved since add three things.

California. The Supreme Court of California wrote to the State Bar on 22 August 2025 asking it to consider moving the 2023 Practical Guidance into the Rules of Professional Conduct and to address agentic AI. COPRAC approved AI-specific comments to six rules on 13 March 2026 (competence, communication, confidentiality, candour, supervision and non-lawyer assistants), with a comment period to 4 May, per the State Bar's public-comment page. The Board of Trustees approved updated Practical Guidance covering generative and agentic AI on 14 May 2026 (Ethics News). The agentic point is the new one: a system that takes actions without a prompt still acts under the lawyer's supervision, and Rules 5.1 to 5.3 apply to what it does.

Alabama. Formal Opinion 2026-01 is the only new numbered state opinion of 2026, according to the legalaicompliance.help tracker. If you practise in Alabama, read it alongside the Alabama Supreme Court's decision of 24 April 2026 dismissing an appeal outright over AI-tainted briefs, which shows how the state's courts treat the same failures.

The rest. Virginia's value-billing opinion and the NYC Bar's opinions on AI recording of client meetings (2025-6, 22 December 2025) and non-client calls (2026-2, 5 August 2026) extend 512 into specific workflows. The state-by-state guide has all seventeen.

Mapping Opinion 512 to a firm policy

The useful exercise is not summarising the opinion but converting it into controls with evidence. This mapping is the skeleton of a compliant policy.

DutyPolicy clauseControlEvidence retained
Competence (1.1)Approved-tool list; mandatory training before accessAccess gated on training completionTraining record with dates and content
Confidentiality (1.6)No client information in consumer accounts; enterprise tools only, with listed data termsAdmin settings: training off, retention set, SSO; DLP on consumer domainsScreenshots of settings; vendor terms on file
Communication (1.4)Standard engagement-letter paragraph; escalation where AI touches strategyTemplate letter; partner sign-off on exceptionsSigned letters; file notes
Candour (3.3)Every citation read in the primary source before filingVerification checklist attached to the filing fileCompleted checklist per filing
Supervision (5.1–5.3)AI output reviewed like a junior's draft; agents run with approval gatesNamed reviewer per matter; agent permissions loggedReview notes; agent logs
Fees (1.5)Bill actual time; subscriptions overhead; pass-throughs disclosedBilling-narrative guidance; pre-bill reviewTime entries; engagement terms

A policy template built on this mapping is shorter than most firms expect, because the duties are old and only the controls are new. Module 6 of the AI Lab for Lawyers course walks through the risk and regulation side of exactly this mapping.

What to do this month

Read the opinion once; it is short. Then audit three things against it: which tools your people actually use and on what terms, whether your engagement letter says anything about AI, and whether anyone verifies citations before a filing goes out. The lawyers sanctioned in 2025 and 2026 did not lack an ethics opinion. They lacked a control at the point where the draft became a filing. Opinion 512 tells you which duty was breached; the sanctions casebook tells you what it cost.

Frequently asked questions

Is ABA Formal Opinion 512 binding on lawyers?

No. The ABA's ethics opinions interpret the Model Rules, which states adopt in their own versions, so 512 binds nobody directly. It matters because state bars, courts and malpractice carriers treat it as the baseline, and as of June 2026 seventeen jurisdictions plus the ABA have formal AI opinions that track its six-duty structure. A lawyer who ignores it will struggle to argue reasonableness.

Do I need client consent to use AI under Opinion 512?

It depends on the tool and the information. Opinion 512 says a lawyer must obtain informed consent before inputting information relating to the representation into a self-learning generative AI tool, because the information could be revealed to others. It adds that boilerplate consent in an engagement letter is not sufficient. For enterprise tools with contractual no-training terms, the duty is to evaluate the tool competently, not necessarily to obtain consent.

Can I bill clients for using AI under the ABA rules?

You may bill the actual time you spend using the tool, refining prompts and checking output. You may not bill for time the tool saved or for time spent learning to use it, unless the client asked for that tool. A general AI subscription is overhead unless the client agrees otherwise up front; per-use charges can be passed through at cost with disclosure. Texas Opinion 705 and California's 2026 guidance say the same.

What is the duty of competence with AI under Model Rule 1.1?

Opinion 512 requires a reasonable understanding of the capabilities and limitations of the specific tool, kept current as the technology changes. It does not require technical expertise. In practice it means knowing that the tool can fabricate authority, knowing how it handles your data, and independently verifying output before relying on it. Texas Opinion 705 phrases it as a reasonable and current understanding of the technology.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Regulation & Ethics

US State Bar AI Ethics Opinions: A 2026 State-by-State Guide

As of June 2026, 17 US jurisdictions plus the ABA have formal AI ethics opinions. None bans AI. They agree on verification, split on client consent and billing.

11 min read

Regulation & Ethics

Billing Clients for AI: What You Can Charge and What You Cannot

Lawyers may bill actual time spent using and checking AI or a fixed fee agreed in advance, but not saved hours, subscriptions or marked-up AI costs. Worked numbers inside.

10 min read

Regulation & Ethics

Do You Have to Tell Clients You Use AI? Disclosure Rules and Practice

When lawyers must disclose AI use to clients under ABA 512, California, CCBE and ÖRAK rules, when consent is required, and a sample engagement-letter clause.

10 min read

Guides

Which AI Tools Train on Your Data? A Confidentiality Checklist for Lawyers

Which AI tools train on your data, plan by plan as of September 2026: ChatGPT, Claude, Gemini, Copilot, legal vendors, plus bar rules and a 10-point checklist.

11 min read

Questions

Is It Ethical for Lawyers to Use AI? The Rules in the US, UK and EU

Yes, every major bar permits lawyers to use AI. It becomes unethical when output is filed unverified or client data is exposed. US, UK and EU rules in 2026.

6 min read

Regulation & Ethics

Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots

Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.

10 min read