What the panels are
A Viewpoints.ai panel is built to mirror the jury pool of a specific U.S. venue. Each simulated juror carries a complete personal profile consistent with that pool: age, education, occupation, income, and the life experiences that shape how a person hears a case. Jurors read your actual case materials, answer questions individually, deliberate as a group when asked, and explain their reasoning in their own words. The same panel can be rerun on a revised version of the case, which makes results comparable across versions.
The published head-to-head
The strongest test of a simulation is running it against real people on the same case, and publishing the result before knowing how it will be received. In a study published in DRI's For The Defense (May 2026), the same case was put to a traditional human mock jury and to a Viewpoints.ai simulated panel. The simulated panel reached the same defense verdict as the human panel and surfaced the same leading themes driving the decision. The full write-up is available as a PDF.
One matched result is evidence, and it is presented as exactly that. It is also, as of this writing, the only head-to-head of its kind published in a legal industry outlet by any vendor in this category.
The benchmark program
Beyond the head-to-head, we run our panels against published, peer-reviewed jury studies whose complete condition-level human results are public. These studies put hundreds of human mock jurors through controlled variations of a case, so they provide something rare: a human answer sheet with known statistics, collected by independent academics with no stake in our results.
We score the simulation on whether it reproduces the human patterns: which condition produces more liability votes, how verdict preferences shift when the evidence changes, and how damages respond to what is asked for. This program is ongoing, and a working paper is in preparation.
What we disclose
Honest validation means publishing the misses along with the hits. Two disclosures matter for anyone relying on simulated panels today:
- Conviction conservatism. In benchmarking against a published criminal study built on a single strand of contested expert testimony, simulated panels convicted less often than the human sample. If you are testing a criminal matter that turns on thin expert evidence, read the simulation's conviction rates as conservative.
- Single numbers are directional. Any one panel's verdict split or damages figure, human or simulated, is a sample. The outputs that hold up best across our validation work are the themes driving votes, the ordering of case versions against each other, and the deliberation dynamics.
The persona engine behind the panels
The jury product runs on the same simulated-persona engine Viewpoints.ai uses for audience research, which has its own replication evidence: an empirical study tested 133 results from human studies and found an 88% replication rate for medium and high effect sizes with matching statistical significance. The engine was built by a Stanford-trained research team working in media psychology and marketing science.
How to use the results
- Trust the themes first. Which arguments move jurors, and which juror profiles they move, is the most reliable and most actionable output.
- Use comparisons, not absolutes. Version A against version B on the same panel is a controlled experiment. A single run is a data point.
- Rerun as the case changes. The panel is repeatable, so treat jury feedback as a continuous instrument rather than a one-time event.
- Confirm the final story however you confirm best. For bet-the-company matters, that may mean a live panel on your best version. The comparison guide covers the sequencing.
"Every vendor in this category claims accuracy. We publish ours, including the parts that are not flattering, because that is what we would demand as buyers."
Common questions on accuracy
How accurate is AI jury simulation?
In the one published head-to-head on the same case, a Viewpoints.ai simulated panel matched the live human mock jury's defense verdict and leading themes. Across benchmarks against published human jury studies, the simulation reproduces human patterns best on themes and on comparisons between case versions, and we disclose the known exceptions on this page.
Can a simulation predict my verdict?
No jury research method predicts a specific jury, and any vendor claiming otherwise should be pressed on evidence. Simulation shows you the distribution of reactions a venue's jury pool produces, the themes driving them, and how the picture changes when your case changes.
What is the simulation validated against?
Three kinds of ground truth: a live human mock jury on the same case, published in DRI's For The Defense; peer-reviewed jury studies with complete condition-level human results; and ongoing internal testing against real tried outcomes.
Where does the simulation read differently from humans?
The clearest known case: on a published criminal study turning on a single strand of contested expert testimony, simulated panels convicted less often than the human sample. We publish differences like this as we find them.
Sources
- DRI, For The Defense, May 2026: human vs. machine mock trial head-to-head.
- Replication study of the persona engine: 133 results from human studies, 88% replication for medium and high effect sizes with matching statistical significance.
- Benchmark details and the working paper will be linked here on publication.