Lindsay Clancy Trial Highlights AI Jury Selection Ethical Risks

The Lindsay Clancy trial jury selection process is the kind of voir dire that makes any shortcut look reckless in hindsight. Over four days in July 2026, Judge William F. Sullivan and the lawyers seated 18 jurors after a process that included group voir dire, a written mental-health questionnaire, and individual sidebar questioning about conditions including postpartum depression, depression, generalized anxiety disorder, and bipolar disorder.[1][2][3] Two seated jurors were excused on the fourth day and replaced before the panel was completed.[3]

That matters because Clancy’s defense had stipulated that she strangled her three children and planned to argue insanity and postpartum psychosis.[1][4] This was not a routine panel where a lawyer could safely reduce people to age, job, neighborhood, family status, or a remembered trial-consultant stereotype. The court had to ask about mental-health experience without turning private histories into crude eligibility screens.

There is no evidence in the research record that AI was used in the Clancy voir dire. The case is useful for a narrower and more important reason: it shows what ABA Formal Opinion 517 would demand if a lawyer tried to use AI-assisted juror ranking in a proceeding where the most relevant questions are also the most sensitive ones.

The Opinion Does Not Let the Lawyer Blame the Tool

ABA Formal Opinion 517, issued July 9, 2025, treats unlawful discriminatory peremptory challenges as outside legitimate advocacy under Model Rule 8.4(g), even when the discriminatory input or recommendation comes from an AI system.[5] The opinion also says lawyers using a juror-selection program must “conduct sufficient due diligence to acquire a general understanding of the methodology employed by the juror selection program.”[5]

That is an ethics opinion, not a statute. Its force depends on the governing jurisdiction, including whether and how that jurisdiction has adopted Rule 8.4(g). But as a professional-responsibility warning, it is hard to miss. If a lawyer exercises a strike and the explanation later has to be given at sidebar, the answer cannot be, “the software ranked this juror as high risk.” The lawyer used the strike. The lawyer owns the reason.

In practical terms, the duty of inquiry starts before the venire walks in. A lawyer using AI to sort prospective jurors needs to know what the program considers, what it claims not to consider, how it handles proxies, whether its scoring can be explained for an individual juror, and what the lawyer will do when the tool’s recommendation conflicts with a lawful, record-ready reason. “General understanding” does not require a trial lawyer to become a machine-learning engineer. It does require enough understanding to avoid outsourcing a discriminatory strike and then calling the outsourcing ignorance.

Why Clancy Is a Stress Test, Not an Accusation

The Clancy panel was selected through a process that deliberately slowed down around mental-health questions. WCVB reported that prospective jurors answered a written questionnaire addressing postpartum depression, depression, generalized anxiety disorder, and bipolar disorder, with individual sidebar examination before Judge Sullivan.[2] MassLive reported that 18 jurors were ultimately seated, 12 women and 6 men.[3]

Those details make the example hard to sanitize. Mental-health experience may matter to a postpartum psychosis defense. But the route from “experience with postpartum depression” to “juror likely to favor the defense” can pass through gender, age, disability, family status, medical history, and assumptions about caregiving. A human lawyer can make the same mistake without AI. The added problem with AI is that the mistake can arrive as a score, wrapped in the confidence of a dashboard.

In a voir dire like this, the safest question is not whether a tool can identify patterns. Many tools can organize information faster than a trial team working from handwritten notes. The harder question is whether the lawyer can separate a legally permissible strike reason from a profile built on variables that merely look neutral because they have been relabeled as “risk,” “leadership,” “persuadability,” or “case fit.”

Due Diligence Has to Reach the Methodology

A defensible AI-assisted jury-selection workflow would leave a paper trail before any contested strike. At minimum, the file should show what tool was used, what the vendor disclosed about its inputs and methodology, what the lawyer asked about prohibited or proxy variables, how the team reviewed the output, and what independent reason supported each strike.

The LACBA analysis of Opinion 517 identifies the structural problem: AI systems can be “a black box even to their developers” and “can autonomously reconfigure their controls.”[5] That does not excuse lawyers from inquiry. It narrows the set of tools a lawyer can safely use without additional verification. If the vendor cannot explain enough for counsel to understand how a juror score is generated, the lawyer has to treat the output as a lead to be checked, not a reason to strike.

The Momus Example Shows the Disclosure Gap

The most useful warning is not abstract. Momus Analytics’ public FAQ says its algorithm “does not take into account race, gender, religion, or country of origin.”[6] Standing alone, that is the kind of sentence a busy trial team might paste into a due-diligence memo and move on.

But AIAAIC, a watchdog repository, documented that a 2020 Momus patent application described consideration of race, education level, and political affiliation in assessing juror “leadership qualities.”[7] That discrepancy is exactly where Opinion 517 becomes operational. The issue is not whether a public FAQ contains the right nouns. The issue is whether counsel asked enough to understand what the system actually does, or did, with variables that may be protected, prohibited, or closely tied to protected status.

The limits of that example matter. The patent material was from 2020. It does not prove the current Momus system works the same way. The research brief identifies no court or regulator that has adjudicated the issue. A lawyer cannot fairly treat the patent reference as a finding of current misconduct. But a lawyer also cannot pretend the discrepancy is irrelevant to due diligence. When public-facing marketing and technical materials appear to point in different directions, the next step is not reliance. It is follow-up.

The follow-up should be specific: whether the current product uses race, gender, religion, national origin, disability, age, political affiliation, education, employment, neighborhood, social-media inferences, purchasing data, or modeled substitutes for any of them; whether any historical training data included such traits; and whether the system can generate a juror-specific explanation that excludes them. If the answer is “proprietary,” the lawyer still has a record to make. Proprietary design is not a litigation privilege against a discrimination challenge.

Chatbot Jury Profiling Makes the Risk Less Contained

The risk is no longer limited to specialized jury-analytics vendors. In a 2026 Florida stalking trial, attorneys Bobby Gonzalez and Valentin Rodriguez reportedly used ChatGPT and Claude to profile jurors and predict outcomes, and they self-reported a 57% success probability.[8] That number should not be treated as an independently validated accuracy rate. Its importance is simpler: lawyers are already experimenting with general-purpose AI tools in jury selection.

Off-the-shelf chatbot use creates a different due-diligence failure point. With a jury vendor, counsel can at least ask for methodology materials, data-input lists, validation information, and contractual representations. With a general-purpose chatbot, the lawyer often cannot meaningfully inspect training data, weighting, embedded stereotypes, or the path from query to output. A polished paragraph about Juror 14’s likely attitudes may feel more explainable than a numeric score, but fluency is not methodology.

That matters most when the lawyer’s question invites profiling. A lawyer who asks a chatbot to infer how a nurse, young mother, veteran, unemployed person, or person from a particular neighborhood will view postpartum psychosis has already built the proxy problem into the request. The model’s answer may only launder the assumption into more comfortable language.

California Shows Where the Standard Can Get Stricter

Batson remains the familiar federal reference point, but some jurisdictions have moved beyond it. California Code of Civil Procedure §231.7, effective January 1, 2026, creates a stricter standard and presumes certain reasons invalid unless clear and convincing evidence shows they are unrelated to a protected group.[5] The LACBA summary identifies neighborhood, lack of employment, dress, attire or personal appearance, and field of employment among the suspect reasons.[5]

Those are precisely the kinds of variables an AI system might find predictive. They are also the kinds of variables that can make a strike explanation sound neutral until a court asks what the variable is standing in for. “Field of employment” may look like life-experience analysis. “Neighborhood” may look like local knowledge. “Appearance” may look like demeanor. Under a stricter proxy-aware rule, the fact that a model treats a factor as predictive does not make it legally safe.

For firms trying cases across jurisdictions, the lesson is not to build one AI-selection protocol around the least demanding forum. A strike file that cannot survive a proxy-focused objection in California may still reveal weak lawyering elsewhere. The jurisdiction changes the burden; it does not change who must answer the judge.

What the Record Should Show Before the Strike

Opinion 517’s duty of inquiry is easiest to respect when the trial team treats AI output as one item in the voir dire file, not as a delegated decision. A useful record would separate three things: what the tool said, what the lawyer independently observed, and what lawful reason actually supported the strike.

• Keep vendor materials, methodology summaries, FAQs, contracts, and any written answers to due-diligence questions in the case file.
• Document whether the tool uses, excludes, infers, or has been trained on protected traits and likely proxies.
• Require juror-specific explanations before relying on any ranking, especially when the case involves mental health, pregnancy, disability, race, religion, age, gender, employment, or neighborhood.
• Record the independent voir dire facts supporting each strike before the strike is exercised.
• Treat chatbot-generated profiles as unverified impressions unless the lawyer can explain the query, inputs, output, and lawful basis for any resulting action.

None of this eliminates judgment. Voir dire has always involved incomplete information, fast decisions, and lawyers reading more into a pause or answer than the transcript later supports. AI can help organize those impressions and force a team to state its assumptions. But once the output starts ranking people for removal, the compliance question becomes sharper: what did counsel know about the model, what should counsel have asked, and can the reason for the strike be defended without leaning on a protected trait or proxy?

That is why the Clancy voir dire is a useful warning even without any evidence of AI use. A panel selected through mental-health questionnaires and sidebar examination is exactly where a model’s appetite for correlation could collide with a lawyer’s duty to make individualized, lawful judgments. In AI-assisted jury selection, the safe harbor is not the vendor demo, the confidence score, or a later claim that no one intended discrimination. It is documented due diligence and independent verification before the strike is made.

References

1. Lindsay Clancy murder trial jury seated, CBS News Boston, link
2. Lindsay Clancy trial jury selection questions July 23, WCVB, link
3. Jury selection for Lindsay Clancy trial completed Thursday afternoon, MassLive, July 2026, link
4. Lindsay Clancy trial could go for weeks and may include a jury home tour, Court TV, link
5. LACBA News, LACBA, link
6. FAQs, Momus Analytics, link
7. Momus Analytics accused of using biased juror selection algorithms, AIAAIC, link
8. AI in Florida jury selection sparks, Yahoo News, link