ABA Formal Opinion 517: What It Requires of AI Jury Selection Tools

The opinion that put AI jury tools inside the ethics rules

In July 2025 the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 517. Most of the coverage framed it as an opinion about peremptory challenges and discrimination. It is that. But it also did something narrower and more consequential for anyone using software in voir dire: it placed AI-assisted juror ranking inside an existing duty, and told lawyers what they are expected to know about the tool they are using.

If your firm uses, or is evaluating, any system that scores or ranks prospective jurors, this is the opinion to read.

What the opinion actually holds

Opinion 517 interprets Model Rule 8.4(g), which makes it professional misconduct to engage in discrimination the lawyer knows or reasonably should know is based on protected characteristics in conduct related to the practice of law. It reads that rule against the constitutional law of jury selection — Batson v. Kentucky and the cases that followed it.

Four holdings matter here.

An unlawful discriminatory strike is not “legitimate advocacy.” Rule 8.4(g) carves out legitimate advice and advocacy. The opinion closes that door for unlawful strikes: conduct declared illegal by courts or a legislature cannot constitute legitimate advocacy. A lawyer who commits a Batson violation has engaged in discrimination that violates the rule.

“Reasonably should know” imposes a duty of inquiry. Where a strike looks discriminatory on its face but a neutral reason has been offered, the lawyer cannot simply accept the neutral reason and proceed. If a reasonably competent lawyer would recognise the stated reasons as pretextual, the lawyer must refrain from the strike.

Whose idea it was does not matter. The opinion works through three scenarios in which the lawyer is not the originator of the discriminatory reason — a client’s directive, a jury consultant’s recommendation, and an AI-assisted juror-ranking tool — and reaches the same place in each. Acting at a client’s direction does not make otherwise unlawful conduct legitimate. Which jurors to strike is a “means” decision belonging to the lawyer under Rule 1.2(a), and Rule 1.4(a)(5) requires the lawyer to consult the client about the legal limits rather than follow an unlawful instruction.

The rule does not extend past the law of peremptory challenges. A strike on a ground substantive law permits — the opinion gives age and marital status as examples courts have allowed — is lawful advocacy and does not violate Rule 8.4(g). Reading the rule more broadly, the opinion says, would create a vague and unworkable standard.

The sentence that matters if you use software

On AI tools, the opinion holds that lawyers should conduct sufficient due diligence to acquire a general understanding of the methodology of the program they are using, citing Formal Opinion 512 on competence with generative AI.

Read that as a practical obligation rather than an abstraction. You do not need to be able to reconstruct the model. You do need to understand the system well enough to answer a specific question: would following this tool’s rankings produce strikes that a court would treat as discriminatory?

That is answerable only if you know what the tool is using as input.

What to ask a vendor before you rely on their rankings

These are the questions the due-diligence standard actually implies:

1. What data does the model consider? Not “public data” — the list. If protected characteristics or close proxies for them are inputs, you need to know that before voir dire, not after a Batson challenge.
2. Can the system explain an individual ranking? A number with no reasoning cannot be inquired into, and the opinion imposes a duty of inquiry. A tool that cannot tell you why it ranked a juror low has made the duty impossible to discharge.
3. Which inputs are proxies? Postcode, neighbourhood, some employment and consumer variables can track race or national origin closely. A model that never sees race can still produce race-correlated output.
4. What is the documented methodology? The opinion’s standard is a general understanding of methodology. If a vendor will not describe it, you cannot meet the standard by using their product.
5. Where does the juror data go, and who else sees it? A separate duty, but it arrives in the same conversation.

A vendor who treats these as hostile questions is telling you something useful.

What this means for jury consultants too

The opinion is not limited to software. A consultant’s recommendation is treated the same way as a client’s directive or a model’s output: if a strike appears discriminatory and the offered reason is one a reasonably competent lawyer would see as pretextual, the lawyer must decline to make it. The name on the recommendation does not transfer the responsibility.

In practice this means the memo you receive from a consultant, and the ranking you receive from a tool, are both inputs to your judgement rather than substitutes for it — and the record of your inquiry is worth keeping.

A working practice for AI-assisted voir dire

• Understand the tool before the trial, not during it. Due diligence conducted under time pressure in the courtroom is not due diligence.
• Treat rankings as hypotheses. A low score is a reason to look more closely at a juror, not a reason to strike one.
• Write down your reason for each strike, in your own words, at the time. If the reason you can articulate is not the reason the tool gave, that gap is worth noticing before a court notices it.
• Know which grounds are lawful in your jurisdiction. The opinion is explicit that the rule tracks the substantive law of peremptory challenges — and that law is not identical everywhere.
• Ask whether the tool’s methodology has changed. Models are updated. Diligence done once against a version you are no longer running is a record, not a safeguard.

Where our own tool stands

VerdictPilot models likely juror response for trial preparation. We think Opinion 517 sets the right expectation, and it is the standard we would want any vendor in this space held to: if a system’s output could shape a strike, the lawyer must be able to see what drove it.

The wider principle is the same one running through the rest of our work — transcript search that returns page and line citations, brief sections attributed to the record. Output a lawyer cannot interrogate is output a lawyer cannot responsibly act on. Opinion 517 makes that a matter of professional conduct in the one setting where the consequences fall on someone who never chose to be there.

Source: ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 517 (9 July 2025), which interprets Model Rule 8.4(g) against Batson v. Kentucky, 476 U.S. 79 (1986) and its progeny, and applies the AI-competence guidance of Formal Opinion 512 (2024). The opinion is published by the ABA and is the authoritative text; this page is a summary and commentary, not legal advice.