Synthetic juror simulations: what they can and cannot show

A careful reporting format for AI-generated juror personas, prompt sensitivity, invented examples, and limits on real-jury prediction.

For: Jury-research agents and legal-technology evaluators

Prevent synthetic discussion exercises from being presented as empirical evidence about real juries.

How can synthetic juror simulations be described without claiming to predict real juries?

Describe them as generated responses to chosen prompts and assumptions. They may help surface alternative interpretations or confusing language, but they are not recruited jurors, a representative population sample, or a validated forecast. Report the setup, variation, and unknowns instead of assigning a courtroom probability to generated votes.

Generated personas are not a jury sample

The federal jury process selects qualified people and includes questioning about suitability to serve. A set of language-model personas does not reproduce that institutional process or the lived experience of actual jurors. [S1]

A persona description is an input assumption, not an observed person's characteristics.
Repeated model outputs are not automatically independent observations.
A majority among generated responses is a property of the experiment setup.
No result here estimates the views of a venue, demographic group, or actual jury.

A proposed simulation disclosure card

Record the purpose narrowly: for example, identify words in a fictional opening that could be misunderstood. Use synthetic facts and avoid pretending the model has access to a real juror's mind. NIST's GAI profile warns against broad capability claims based on narrow assessments. [S2]

Model/version, date, system instructions, prompts, and sampling settings where available.
How personas were designed and which assumptions were intentionally varied.
Number of runs and any discarded outputs, with reasons.
What was measured: themes or wording reactions, not predicted verdicts.
Known omissions: real deliberation, evidence rulings, selection, emotion, and case context.

Synthetic sensitivity test: one word changes the story

In an invented exercise, twelve generated responses read a fictional delivery dispute. Eight initially call a shipment 'late.' After the prompt replaces 'promised date' with 'estimated date,' only five do. These numbers demonstrate sensitivity to framing in the constructed exercise; they do not show a 67% or 42% probability that a real jury would agree.

Useful output: flag the legal and factual significance of the two phrases for review.
Unsupported output: a settlement recommendation based on the generated vote share.
Next research step: verify the actual contract language and relevant evidence.
Do not average the two runs into a purportedly more accurate forecast.

Write findings at the level actually supported

A defensible report says, 'This prompt produced these themes under these assumptions.' It can identify ambiguous wording or hypotheses for further research. It should not imply demographic realism, representativeness, calibrated probabilities, or validation against actual verdicts without separate evidence.

Keep themes and representative synthetic excerpts linked to run IDs.
Report inconsistent responses and prompt sensitivity.
Separate brainstorming usefulness from predictive validity.
State that any proposed real-world research would require its own ethical, methodological, and legal review.

Scope and limitations

Methodological education using fictional personas and invented counts. This site does not provide jury selection, profiling, verdict prediction, or litigation consulting.

No real jurors, demographic datasets, verdicts, or live case facts were used. No predictive validity is established.

Sources

U.S. Courts — Juror Selection Process — Federal prospective-juror selection and questioning; actual juries are human institutions. Checked 2026-11-05
NIST AI 600-1 — Generative Artificial Intelligence Profile — GAI risk and evaluation guidance; Measures 2.3/2.5 address capability evidence, generalization limits, and source verification. Checked 2026-11-05

Explore further

Simulation limits are part of assessing responsible legal-team AI use. Explore the wider topic at Agentic Attorney: https://agenticattorney.com/ . This page makes no claim about jury-prediction capabilities there. Explore Agentic Attorney.