The client was IBM, defending an antitrust case brought by California Computer Products. The professor was Donald Vinson, who taught marketing at the University of Southern California and had spent his career studying how ordinary people form opinions about things they do not fully understand.
His idea was simple to describe and expensive to run. Once the real jury was seated, he recruited a second group of people — matched to the real panel on age, occupation, background — and sat them in the public gallery. They watched the same trial, entered and left when the jury did, and heard nothing the jury did not hear. Each evening he debriefed them by telephone and sent the findings to IBM’s trial team, who used them overnight.
He called it a shadow jury. Fifty years later, trial teams with the budget for it still do exactly this, and the firms Vinson went on to found — Litigation Sciences, then DecisionQuest in 1989 — are still in business.
The practice is worth understanding for two reasons. It is the clearest working answer anyone has produced to the question what is my jury actually thinking? And its history contains a fact the industry rarely volunteers.
A modern shadow panel runs four to eight people. They are recruited to mirror the empanelled jury as closely as the consultant can manage — not just demographically, but on attitudes relevant to the case: how they feel about corporations, about police, about damages awards, about personal responsibility.
They sit in the public part of the courtroom. They do not know the trial team. They must never speak to a real juror — that is not a matter of etiquette but of professional conduct, since Model Rule 3.5 forbids a lawyer from communicating with a juror during the proceeding, and a surrogate acting on counsel’s behalf falls squarely within the concern.
Each evening a consultant debriefs them: what did you understand, what confused you, which witness did you believe, where do you stand now. By morning the trial team has a report describing how the day landed on people who resemble the twelve who will decide it. Cross-examinations get retooled. Witness order changes. Closing themes get rewritten. Settlement posture shifts.
Since 2022 there are remote versions, where surrogates watch a same-day video feed instead of sitting in the room. The cost profile is similar.
Published figures for adjacent services give the shape of it. A mock trial — a single exercise, typically one day — runs $10,000 to $60,000 and up, scaling with panel size and venue. General jury-consulting engagements run $5,000 to $25,000 per trial, with day rates of $3,500 to $7,500; high-profile matters exceed $100,000. Mock jurors themselves are paid $150 to $500 a day.
A shadow jury costs more than a mock trial, for a structural reason: it is billed across the entire length of the trial. Six to eight people every day, plus a consultant every evening, plus a written report every night. A two-week trial is ten of those days.
Which means the tool that best answers what is my jury thinking is available, in practice, to insurers, corporate defendants, and plaintiff firms working large contingency cases. It is not available to a public defender’s office, to appointed counsel, or to a solo practitioner with a client who is paying by the month.
That asymmetry is not incidental to the criminal justice system. It is a description of it.
Here is what surprised us when we went looking.
After fifty years of practice, there is no peer-reviewed published study establishing that shadow juries predict real verdicts.
We looked for one. What exists is a 2025 doctoral dissertation from Chapman University — a content analysis of 82 civil trials across 40 states, comparing parallel-jury findings against the verdicts that followed. Its results are genuinely encouraging: liability findings tracked the real outcome in 88% of cases, and damages awards fell within a 25% band in 70%.
That is real work and we do not want to diminish it. But it is a dissertation rather than a peer-reviewed article, it is one year old, and nobody has replicated it. Everything else offered as evidence is a vendor recounting a case that went well.
We say this as people building a competing approach, so read it with that in mind. But the observation cuts in an unexpected direction: an established, expensive, widely respected practice has been sold for five decades on professional judgement and reputation rather than on measured accuracy. Anyone entering this field — us included — should be honest that the bar for evidence here has been set remarkably low, and should not mistake tradition for validation.
Running alongside the history of jury science is a second history, about whether a court will accept a result a computer helped produce. It is worth knowing because it is the closest thing lawyers have to settled ground.
In February 2012, Magistrate Judge Andrew Peck of the Southern District of New York wrote an opinion in Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182, containing this sentence: “This Opinion appears to be the first in which a Court has approved of the use of computer-assisted review.” The dispute was about reviewing roughly three million documents. The defence proposed predictive coding rather than human review of everything.
Peck approved it, and his reasoning leaned on empirical work rather than intuition — specifically Maura Grossman and Gordon Cormack’s 2011 study in the Richmond Journal of Law & Technology, which compared machine-assisted review against exhaustive manual review on the same document sets. The machine achieved 76.7% recall and 84.7% precision. The human teams achieved 59.3% recall and 31.7% precision. Human reviewers, depending on the topic, missed between 20% and 75% of the relevant documents.
Three years later, in Rio Tinto v. Vale, 306 F.R.D. 125 (2015), the same judge wrote that using this technology had become “black letter law” — a party no longer needed permission. And in Hyles v. City of New York (2016) he declined to force a party to use it, holding that the producing party chooses its own method.
There is a pleasing detail here. The standard practitioner text on this subject, Perspectives on Predictive Coding and Other Advanced Search Methods for the Legal Practitioner (ABA, 2016), was co-edited by Ralph Losey — who was defence counsel in Da Silva Moore, the lawyer who actually ran the technology in the case that made the law — and carries a foreword by Judge Peck himself.
It is important to be precise, because this authority gets stretched.
The trilogy is about discovery method: whether it is reasonable to use an algorithm to find documents, judged the way any other search method is judged. It says nothing about whether artificial intelligence assisted in drafting a brief, and it never has — that question did not meaningfully exist until 2023.
That newer question is governed separately and unevenly, by individual judges’ standing orders, which differ from courtroom to courtroom. We have written about two in the District of Colorado that were entered fourteen months apart and do not require the same thing.
Worth noting for Colorado practitioners: on 8 January 2026 Colorado became the first state to adopt amendments to its Rules of Professional Conduct addressing artificial intelligence directly. Notably, they impose no new disclosure obligation. They re-anchor the duties that already existed — competence, candour to the tribunal, confidentiality, supervision — to the fact that these tools are now in use.
Vinson’s insight was not technological. It was that a jury is a group of specific people receiving a specific story, and that you can learn something real by watching people like them receive it. Everything since — the focus groups, the mock trials, the scoring models, our own work — is a variation on that observation.
What has changed is not the insight but its price. A method that requires eight paid surrogates and a consultant for every day of a trial will always belong to the side that can afford it.
That is the problem we are actually working on. Not replacing a trial lawyer’s judgement, which no software should attempt, but making the thing that informs it cost something other than fifty thousand dollars.
Sources: Kelly Anthony, “The Parallel Jury: Describing Its Processes and Procedures and Evaluating Its Effectiveness in Predicting Jury Verdicts” (Chapman University, 2025), DOI 10.36837/chapman.000639. Maura R. Grossman & Gordon V. Cormack, “Technology-Assisted Review in E-Discovery Can Be More Effective and More Efficient Than Exhaustive Manual Review,” 17 Rich. J.L. & Tech. 11 (2011). Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012); Rio Tinto Plc v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015); Hyles v. City of New York, 2016 WL 4077114 (S.D.N.Y. 2016). Cost figures from published guidance by IMS Legal Strategies and the National Law Review. This page is commentary and general information, not legal advice.