Synthetic respondents are the Wild West. It is time for a standard.

Synthetic respondents went from research curiosity to line item in about two years. Dozens of companies, ours included, now sell AI that answers surveys and interviews in place of people. Buyers are being asked to make real decisions on the output, and there is still no shared way to say whether that output is any good.

Three signs of a market without rules

The words do not mean anything fixed. The same product is sold as digital twins, synthetic respondents, synthetic users, AI personas or simulated audiences, and the same word covers very different things: a model of one real person, or a paragraph describing a segment. Our post on the terms walks through the differences.

Accuracy claims are not comparable. One vendor reports parity with real respondents, another a confidence score, another a correlation, another a share of test-retest reliability. Almost none are independently audited, and most do not say what the number is measured against.

Even the research disagrees, because it measures different things. A Stanford study found twins built from two-hour interviews reached 85% of people's own test-retest consistency on standard survey questions. A Columbia mega-study of twins built from 500 answers per person found an average correlation of 0.20 with real responses on new tasks. Both are careful work. Without a standard, a vendor can quote whichever one suits it.

The industry has started

The research associations have moved first. ESOMAR published questions for buyers of AI services in 2024 and guidance for buyers of synthetic data in 2025, including a split between boosted data, fully synthetic data and synthetic personas. The Insights Association said in 2026 that synthetic participants should be fully disclosed, that reports should say whether findings rest on people, synthetic participants or both, and that synthetic data should not be presented as equivalent to human responses without validation.

A newer effort, the Synthetic Data Standard, goes further and proposes independent certification. Its SDX-1 standard assesses vendors across six domains: grounding, accuracy, fairness, freshness, transparency and responsibility, with transparency covering honest labelling, disclosed limitations and traceability. It describes itself as an independent certification body shaped by an advisory council of insights leaders and academics, and has published a first field guide on the types of synthetic data.

What a standard should require

Whoever ends up setting it, a useful standard would make a few things non-negotiable.

• Grounding disclosed. Say what each synthetic respondent is built from: a real person's interview, survey history, or a description.
• Accuracy against a baseline. Report results relative to random guessing and to people's own test-retest consistency, on data the model was not built from.
• Distributions, not averages. Report the spread and the subgroups, because averages hide the failures.
• Fairness by segment. Publish the gap between the best and worst served groups.
• Freshness. Say when the grounding data was collected, and refresh it.
• Labelling in every report. Mark which findings rest on people and which on synthetic respondents.
• Independent checks. Let a third party rerun the evaluation.

What we do now

Jean grounds each digital twin in a real person's interview, runs real interviews alongside twins in the same study, and publishes evaluation results on its open evaluations page. We would rather be held to a public bar than set our own, and we will support whichever standard the industry converges on.