Child and Teen Audiences on a Synthetic Panel

Methodology · 10 min read

TL;DR: Under-18 audiences are the one cell where a synthetic panel gets easier and less trustworthy at the same time. The ICC/Esomar Code defines a child as 12 and under, and its duties attach to a human being, which a persona is not, so parental consent and safeguarding have nothing to attach to. The grounding weakens in the same direction: the US Census Bureau measured a 4.6 percent net undercount of children aged 0 to 4 in 2010 against a net overcount of adults. And a model barely steers by the age it is given. Use the panel for the adult gatekeeper and to rehearse the instrument. Keep the child readings on regulated fieldwork.

Can you use a synthetic panel to research children and teenagers?

**Not for readings about the children themselves. A census-grounded synthetic panel can represent the adults around a child and can rehearse a youth questionnaire before fieldwork. It cannot stand in for a child respondent, because the population totals are weakest at the youngest ages and a model barely steers by the age it is given.**

Youth research is the most governed corner of consumer insight. Reaching a 9-year-old means parental consent, a vetted moderator, a shortened session and a signed-off topic list. A synthetic panel removes all of it in one step.

Two things change when an under-18 study moves onto personas, and they do not cancel out. The rules that make youth fieldwork slow stop applying. The evidence that makes it worth doing gets thinner at the same time. The result is the freest cell in your panel attached to the least defensible row in your deck.

Geography has the same shape, and this blog has covered it: where a census-grounded panel stops going local. Age is harder, because age carries a legal regime that a postcode does not.

Which research rules stop applying when the respondent is synthetic?

**Almost all of the ones that make youth fieldwork slow. The ICC/Esomar Code defines an individual as a human being, specifically to distinguish one from a synthetic persona, and its child provisions bind researchers collecting data from children as data subjects. With no child in the study, parental consent and age verification have nothing to attach to.**

The 2025 revision of the ICC/Esomar International Code is unusually direct. Its definitions state that individual or person "refers to a human being to differentiate from a synthetic, virtual/digitally created persona or entity".¹ Article 2 then sets out the child provisions. Consent "from a parent, legal guardian or responsible adult must be obtained whenever children are involved as data subjects". Researchers must establish a data subject's age before collecting further personal data from a child, and they "must exercise special care when considering the involvement of children and young people in research".¹

Each of those duties presupposes a person. A persona has no parent to ask and no age to verify.

The boundary itself moves between codes, which matters across markets. ICC/Esomar defines a child as "a person aged 12 years and under" and a young person as "an adolescent aged 13 to 17".¹ The MRS draws one line instead of two: "Under the Code, children are defined as those aged under 16 years", with Rules 16 to 22 carrying the mandatory provisions.² A 14-year-old is a young person to one body and a child to the other.

What remains when nobody answers is narrower and entirely yours: disclose that the data is synthetic, and stay accurate about what the age cell represents. Those duties are covered separately in which standards duties apply to a synthetic panel.

Why is census grounding weakest for the youngest ages?

**Because young children are the hardest people for a census to count. The US Census Bureau measured a 4.6 percent net undercount of children aged 0 to 4 in the 2010 Census, 2.2 percent for ages 5 to 9, and a 0.5 percent net overcount for ages 10 to 17. Adults were overcounted by 0.7 percent.**

A census-grounded panel inherits the quality of the totals it is calibrated against, and those totals are not uniformly good by age.

The US Census Bureau analyzed this. Its 2017 report on the 2010 Census found that "only the two youngest age groups had a net undercount in 2010. The youngest children had a net undercount of 4.6 percent while children age 5 to 9 had a net undercount of 2.2 percent. Children age 10 to 17 had a net overcount of 0.5 percent".³ Across three decades the gap widened: young children went from 1.4 percent undercounted in 1980 to 4.6 percent in 2010, while adults moved to a 0.7 percent overcount.³

The reasons are about who fills in the form. The report finds that "children who were not related to the householder had the highest CFU add rates" and that householders "make errors by omitting young children who were grandchildren and other relatives".³

So the error is structured, not random. The children a census misses most live in the least standard households, which is often the variation a youth brief most wants to see.

How much does a persona's age actually change its answers?

**Less than the data suggests. A 2026 arXiv study of demographic identity in a language model found an 11-point real gap between the 18 to 29 and 50 to 64 brackets moved by only 1.7 points under a full persona swap, about 15 percent of the true difference, even though age was among the most strongly encoded attributes tested.**

The second limit sits in the model rather than the data, and the evidence comes from adults.

Fathin Difa Robbani's 2026 study separates three properties of a demographic attribute inside a language model: whether it can be read out, whether it is arranged faithfully, and whether the model steers by it. Age scored well on the first two: held-out probes recovered age-by-party orderings at a Spearman rho of 0.59 to 0.63, among the strongest attribute types tested.⁴ Steering was the weak link. An "11-point real gap" between the 18 to 29 and 50 to 64 brackets "moves by only 1.7 points under a full persona swap", roughly 15 percent of the true difference, and the paper's summary of the pattern is blunt: "causal use does not track fidelity".⁴

A second 2026 paper agrees from another direction. Taday Morocho and colleagues conditioned two open-weight models on personas built from World Values Survey respondents and scored 31 items against those respondents' real answers. Persona prompting "does not yield a clear aggregate improvement" and "in many cases, significantly degrades performance", with distortions concentrated in "a small subset of questions and underrepresented subgroups".⁵

Read the two together. A model holds a usable internal map of age and then mostly declines to drive with it. If a 30-year gap between adults moves an answer by under two points, a prompt that says "you are 11" has no measured claim on an 11-year-old's answer. Nobody has tested whether it does, because there is no benchmark survey of 11-year-olds to test against. That absence is the finding.

The tiering in which low-incidence audiences a census-grounded panel can represent applies here with one correction. Under-18s are not rare. They are well documented as a population and barely as respondents.

What can a synthetic panel do on a youth brief?

**Three jobs, all of them adjacent to the child rather than about the child. Represent the adult who decides and pays. Rehearse the questionnaire before a regulated field, which is where most youth studies lose time. And screen a long list of concepts down to the few worth putting in front of young people.**

Three jobs, sorted by who the respondent has to be.

One. The gatekeeper. In most under-13 categories the budget, the permission and the veto sit with a parent, and a parent is an ordinary adult respondent on a census-grounded panel. A study of what that parent will allow, pay for and refuse is not a youth study with a workaround in it. It is an adult study that was labeled wrongly.

Two. The rehearsal. Youth fieldwork is slow to re-field, so a defect found in the first wave is expensive. An adult panel will not tell you whether a 9-year-old understands your wording, which needs cognitive testing with children. It will catch the defects that have nothing to do with age: dead routing, double-barreled items, a scale that collapses, a grid nobody finishes. The protocol in the multi-market pretesting playbook carries over unchanged.

Three. The screen. Fourteen concepts cannot all go into a session with children. Ranking them on an adult panel costs an afternoon at one credit per persona per question on PersonaHive, and what comes back is a shortlist, not a result.

None of the three produces a number about children. If the deliverable has a row reporting what 13 to 17 year olds think, that row needs 13 to 17 year olds in it.

Frequently asked questions about synthetic panels and under-18 audiences

**Is it legal to simulate a child respondent?**

Nothing in the research codes prohibits it, because their duties attach to a human data subject.¹ Legality is not the binding constraint here. Validity is.

**Can I report a synthetic result for a 13 to 17 segment?**

Only with the label attached. State that the cell is simulated and what it was calibrated against, and do not place it beside a human-sourced row as though the two are the same measurement.

**Does a richer persona brief fix the age problem?**

No. The 2026 study's finding is that the model under-uses an age it already represents well, so adding prompt detail does not supply the steering that was missing.⁴

**Do the census undercount figures apply outside the United States?**

The 4.6 percent figure is specific to the 2010 US Census.³ The size of the error is country-specific, so check the national statistics office for each market you run.

**What about 16 and 17 year olds, who are adults under some codes?**

The legal position gets easier and the measurement position does not. The MRS line at under 16 puts a 17-year-old outside its definition of a child,² but the model evidence on age conditioning is unchanged.

What should you run before commissioning youth fieldwork?

**One thing: the questionnaire, on the adult cell, before anyone books a moderator. It is the cheapest step in the project and it removes the defects that cost the most to fix once children are in the room. Treat the output as a defect list and a shortlist, never as a youth finding.**

Take the questionnaire or discussion guide you intend to field with children. Field it on an adult panel first and read it for two things only: which items break, and which concepts survive. Then commission the regulated study on what is left.

You can run that pass on a free PersonaHive account with the credits included, at one credit per persona per question. Defining the audience and building the questionnaire by hand are free.

Keep the youth readings on fieldwork with real young people, consented and supervised. That is not a limitation of one platform. It is where the evidence is.

Sources

• ICC/Esomar International Code on Market, Opinion and Social Research and Data Analytics (2025) — ICC and Esomar, International Chamber of Commerce and Esomar
• Guidelines for Conducting Data Collection Activities with Children — Market Research Society, Market Research Society
• Investigating the 2010 Undercount of Young Children — Decennial Statistics Studies Division, US Census Bureau
• Readable, Faithful, Used: Three Dissociable Properties of Demographic Identity in a Language Model — Fathin Difa Robbani, arXiv
• Assessing the Reliability of Persona-Conditioned LLMs as Synthetic Survey Respondents — Taday Morocho, Cima, Fagni, Avvenuti and Cresci, arXiv, published in Companion Proceedings of the ACM Web Conference 2026