Using AI in a VA Practice: ABA Formal Opinion 512 for Representatives

ABA Formal Opinion 512, issued July 29, 2024 by the ABA Standing Committee on Ethics and Professional Responsibility, applies the existing Model Rules of Professional Conduct to generative AI tools. In short: understand the tool well enough to use it competently and check its output; protect client information and get informed consent where the tool could expose it; tell clients when the facts call for it; never file unverified output; supervise staff and vendors; and charge only reasonable fees. For a VA practice, the same duties sit beside VA's own standards of conduct in 38 C.F.R. § 14.632.

This is a plain-language summary for representatives, not legal advice and not an ethics opinion. ABA formal opinions are advisory and do not bind any court or bar; your state's rules and its own guidance control. Read the opinion itself, on the ABA's website, before relying on any point here. The author is not a VA-accredited representative.

Who this applies to in VA practice

• Attorneys. Accredited attorneys are bound by VA's standards of conduct and, in addition, may not engage in conduct prohibited by the professional conduct rules of any jurisdiction where they are licensed (§ 14.632(d)). Most states' rules follow the Model Rules, so Opinion 512 is directly relevant.
• Claims agents. Agents are not bound by a state bar's rules, but § 14.632 requires competent representation and diligence, and bars disclosing information VA provided for representation without the claimant's authorization (§ 14.632(b) and (c)). The opinion is a useful, careful reading of what those same duties mean when a tool is involved.
• VSO representatives. The standards of conduct in § 14.632 apply to representatives as well, and your organization may have its own policy. Follow it.

What the opinion says, rule by rule

Competence (Model Rule 1.1)

You do not need to be an AI expert, but you need a reasonable understanding of what the tool you use can and cannot do, and to keep up as it changes. Generative tools can produce confident, plausible text with no basis in fact. The opinion says relying on output without an appropriate degree of independent verification or review can fall short of competence. How much checking is appropriate depends on the tool and the task, and the lawyer stays responsible for the work.

In a VA file, that means checking every fact against the record page it came from and every citation against the regulation's text. Section 14.632(b) defines competent representation as including an understanding of "the applicable provisions of title 38, United States Code, and title 38, Code of Federal Regulations." A tool's summary of a regulation is not a substitute for reading it.

Confidentiality (Model Rule 1.6)

Before entering client information into a tool, assess the risk that it could be disclosed or accessed, inside or outside the firm. For self-learning tools, where what you enter could surface for others, the opinion says the client's informed consent is required first, and that general boilerplate in an engagement letter is not enough. If no client information goes in, no consent is needed.

Published analyses of the opinion describe a baseline step: read and understand the tool's terms of use, privacy policy, and related contract terms, or rely on a colleague or outside expert who has, to learn who can access what you enter, how long it is kept, and whether it is used for training.

In a VA file, client information includes medical records, Social Security numbers, and everything VA disclosed to you under the 21-22a. Section 14.632(c) separately prohibits disclosing information VA provided for representation without the claimant's authorization.

Communication (Model Rule 1.4)

The opinion does not require telling every client about every tool. It says disclosure depends on the facts, including when the client asks, when the engagement agreement requires it, when informed consent is needed under Rule 1.6, when the tool affects the fee, and when its output will significantly influence a decision in the matter.

Candor and meritorious claims (Model Rules 3.1, 3.3, and 8.4(c))

Output has to be reviewed before it is used, including checking for citations to authority that does not exist and for misstatements of law or fact. The opinion notes that even an unintentional misstatement to a tribunal can be a misrepresentation.

In a VA file, the parallel duties are the requirement to be truthful in dealings with claimants and VA (§ 14.632(a)(2)) and the bar on conduct involving misrepresentation (§ 14.632(c)). A rebuttal that misquotes a rating criterion or invents a precedent is your misstatement, not the tool's.

Supervision (Model Rules 5.1 and 5.3)

Lawyers with managerial authority should set clear policies on how the firm may use these tools, and supervisors should make sure lawyers and staff are trained in using them ethically and practically. The duties that apply to outside vendors, which earlier opinions worked out for cloud computing and outsourcing, apply to AI providers too.

Fees (Model Rule 1.5)

A lawyer billing hourly may bill only for time actually spent, not for time the tool saved. The opinion also says a lawyer generally may not charge a client for time spent learning a tool the lawyer will use across many matters; that a tool that works like ordinary office equipment is overhead, not a billable cost; and that a flat fee that ignores large efficiencies may be unreasonable.

In a VA file, most fees are contingent percentages of past-due benefits under § 14.636, which must still be reasonable. The reasonableness factors in § 14.636(e) include the time spent and the work done, so a fee that made sense for a month of manual file review should be measured against the work actually required.

A responsible-use checklist for a VA office

1. 1Write a one-page policy. Which tools are approved, for which tasks, and what may never go into an unapproved tool.
2. 2Vet each vendor in writing. Where data is stored, who can access it, whether it is used to train any model, how long the vendor and its AI providers keep it, what happens on a breach, and how data is deleted when you leave.
3. 3Decide when you need consent. If a tool could expose client information, get the client's informed consent in a specific, plain explanation, not a boilerplate clause.
4. 4Verify everything that leaves the office. Every fact against its record page, every citation against the regulation or decision, every number against its source.
5. 5Keep the judgment human. The tool drafts and organizes; the representative chooses the lane, makes the argument, and signs.
6. 6Train and supervise. Show staff the policy and how to check output, and spot-check their work.
7. 7Bill honestly. Do not bill hours the tool saved, and do not charge clients for learning it.
8. 8Record what you did. A note in the case diary of what was drafted with a tool and how it was checked answers questions later.

Frequently asked questions

What is ABA Formal Opinion 512?

It is the ABA Standing Committee on Ethics and Professional Responsibility's opinion on lawyers' use of generative AI tools, issued July 29, 2024. It explains how existing Model Rules on competence, confidentiality, communication, candor, supervision, and fees apply to those tools. It is advisory, and each state's own rules and guidance control.

Does ABA Formal Opinion 512 apply to VA-accredited claims agents?

Not directly; it interprets rules for lawyers. Agents are bound by VA's standards of conduct in 38 C.F.R. § 14.632, which require competent and diligent representation and bar disclosing information VA provided without the claimant's authorization. The opinion's reasoning is a careful guide to what those duties mean when a tool is involved.

Do I need a client's consent to use AI on their VA claim?

It depends on the tool and what goes into it. The opinion says that for self-learning tools that could disclose information relating to the representation, informed consent is required before entering that information, and boilerplate in an engagement letter is not enough. If no client information is entered, consent is not needed. Check your state's guidance too.

Can a representative file an AI-drafted VA rebuttal?

Only after reviewing and verifying it. The representative is responsible for every fact, citation, and argument, and must be truthful in dealings with VA (38 C.F.R. § 14.632). Check each statement against the record and each citation against the regulation's text before signing.

What should I ask an AI vendor before using it on client files?

Where the data is stored, who can access it, whether it is used to train any model, how long the vendor and its AI providers keep it, what happens on a breach, and how it is deleted when you leave. Get the answers in writing and read the terms of use and privacy policy yourself.