What an AI panel is, and what it is not.

The method

A target is defined as a set of normalised axes. From those axes the platform computes a quota matrix in code, then asks a model to fill each cell — so a panel spans the declared ranges instead of producing eight variations of the same modal person. Each persona carries deliberate internal contradictions, because respondents without tension answer everything agreeably, which is precisely what real respondents do not do.

Each persona then answers the same version-locked question battery, one call per turn, producing numeric fields (purchase intent, problem severity, comprehension, willingness to pay, sentiment) and objection codes from a fixed taxonomy at generation time.

By default personas do not see each other’s answers. A “group dynamics” toggle can show each persona sampled prior answers, which reproduces herding — a known artefact of real focus groups. Those are two different research instruments and the report states which one ran.

How synthetic output is kept labelled

• Database. A response row must be either human or synthetic, and a check constraint makes an unlabelled synthetic row impossible to insert. A persona record cannot be marked non-synthetic.
• API. Every read path returns the source alongside the text. There is no endpoint that returns response text without it.
• Interface. Synthetic sessions carry a distinct accent colour and a persistent chip in the session chrome — not a footnote. Simulated respondents are called “simulated personas”, never “participants”.
• Export. The header band is written server-side into every PDF and CSV, and every CSV row carries a source column. A CSS watermark that a screenshot can defeat is not a control.
• Statistics. Synthetic results are never reported as percentages of a population. Until human comparison data exists they are shown with an explicit “uncalibrated” banner rather than an implied accuracy.
• Separation. Synthetic sessions are blocked from the participant-facing surface entirely.

Known limitations

• A model cannot tell you what people do — only what a described person plausibly says.
• Personas are drawn from training data, so they inherit its blind spots. They are worst exactly where a segment is under-represented online, which is often the segment worth researching.
• Small panels produce wide error bars. Panel counts are always shown before percentages, because “75% of eight people” is six people and the percentage implies precision that is not there.
• An AI panel cannot be used as human-subject research and may not be represented as such.

Calibration — the honest version of an error bar

The same battery is designed to run in live human sessions. Every live session produces a labelled pair of synthetic and human means for a segment, accumulating into per-category calibration figures — statements like “synthetic panels overstate purchase intent by 1.8 points in this category, n=340 human responses”.

None of this exists yet. There is no human data, so there is no calibration, and any category would display as uncalibrated. Saying so is the point: in a category full of overclaiming, a published error bar is the only honest differentiator, and it has to be earned rather than asserted.