How to build digital twins that behave like real people

A digital twin is easy to demo and hard to build well. Write a persona, hand it to a chat model, and it will answer anything. Whether those answers resemble the person is decided long before the demo, in a handful of stages most vendors do not describe. This is how we think about them.

1. 01RecruitThe right people, with consent
2. 02InterviewDeep, adaptive, in their words
3. 03StructureWhat they said, why, what they did
4. 04ModelA model trained on human behavior
5. 05ValidateAgainst held-out real answers
6. 06RefreshRe-interview as people change

1. Recruit the people you actually care about

A twin can only resemble someone who was in the data. If your buyers are a 2% segment in three regions, those people need to be interviewed, not approximated from the general population. Research consistently finds language models are least accurate for groups far from their training data; one study measured a correlation of minus 0.70 between a country's cultural distance from the US and how closely model answers resembled its people. Recruiting is the first quality decision, and consent to be modeled belongs in it.

2. Interview deeply, not widely

The grounding data matters more than anything else. In the Stanford study of 1,052 people, twins built from two-hour interviews reached 85% of people's own test-retest consistency on the General Social Survey. Twins built from demographics alone reached 71%, and from a written persona 70%. The interview was the difference.

Two practical points follow. A good interviewer follows up: it asks why, and it asks about what the person did, not only what they think. And depth beats length past a point: even with 80% of the interview removed, the Stanford twins still reached 79%.

3. Structure what people said, why, and what they did

A transcript is not yet a twin. We break each interview into three things the twin has to hold: what the person says about a topic, the reasons they give, and what they actually did. The third matters most. People are better at reporting what they did than predicting what they will do, and stated intentions are where synthetic answers drift toward the agreeable.

4. Run twins on a model trained for people, not for helpfulness

General assistant models are tuned to be helpful and agreeable, and that tuning works against simulation. The SimBench benchmark found instruction tuning improves answers on questions where people agree and hurts them where people disagree, which is exactly where research is useful. Other studies find assistant models give answers that are too positive and too uniform.

Two things help. Ground every answer in the person's own interview rather than a summary of their demographics. And use a model trained on human behavior: what people say, what they do, and why, rather than one trained to be a good assistant. Research on willingness to pay found models performed best when tuned on prior human data from the same category. A foundation model of human behavior is the general version of that idea.

5. Validate against people before anyone uses the twins

Before twins answer a new question, check them on old ones. Hold out answers the twins were not built from, compare whole distributions rather than averages, look at the worst served segment, and report the result as a share of people's own consistency. Our post on evaluating synthetic respondents lists the checks.

6. Keep them fresh

People change: they switch banks, have children, lose jobs, change their minds. Models change too; one study found the same prompts gave different survey results three months apart. ESOMAR notes that time-sensitive measures such as advertising awareness are harder to synthesize than stable ones. Twins need a date on their grounding data and a schedule for re-interviewing, and every new study with real people is a chance to refresh the twins it ran beside.

What this means for buyers

Ask every vendor the same six questions, one per stage. Who was interviewed and did they consent? How deep were the interviews? What does each twin hold? What model runs them? How were they validated, against what ceiling? When was the grounding data collected, and when will it be refreshed?