Synthetic respondents are AI-generated answers that stand in for real survey participants. According to a February 2026 ESOMAR paper co-authored by the synthetic data vendor Fairgen and Google, they are valid only for boosting close-ended quantitative data on top of a real sample of at least 300 people, are not appropriate as a substitute for statistical inference, and are poor at surfacing novel findings. Qualitative and open-ended research remains the domain of real human research.
Three different things called "synthetic"
The phrase covers three distinct techniques, and conflating them is where most confusion starts.
Synthetic augmentation takes a real survey and generates additional records to boost small subgroups, so a base of 40 respondents in a segment can be analysed as if it were 200. Digital twins build a model of a known population from its historical data and query the model instead of the people. LLM-based synthetic respondents ask a large language model to answer a questionnaire as if it were a member of the target audience.
What the vendors' own research concludes
In February 2026 Fairgen, a synthetic data company, and Google published a paper through ESOMAR evaluating these approaches. Because the authors sell the technology, the limits they concede carry weight.
| Approach | Where it holds | Where it fails |
|---|---|---|
| Augmentation | Close-ended quantitative data; real base of n at least 300 | Open-ended, qualitative, small or novel populations |
| Digital twins | Stable, well-documented populations | Reproduces every bias in the training data; blind to change |
| LLM respondents | Directional exploration, questionnaire testing | Statistical inference, novel findings, anything high-stakes |
The paper's language on LLM respondents is direct: they are "not appropriate as a substitute for statistical inference". On qualitative work it is more direct still: open-ended and qualitative research "remain firmly within the domain of real human research". Primary fieldwork remains the gold standard.
The novelty problem
The most important limitation is not statistical. A model trained on the past can only recombine what it has seen. If your question is "what do buyers think of a product category that existed last year", a synthetic answer will be plausible. If your question is "what has changed since the regulation passed in March", or "why did our win rate fall last quarter", the model has no way to know, and it will produce a confident answer anyway. Decisions that matter are almost always about what is new.
Why synthetic is winning anyway
Synthetic data is popular because real online data has become unreliable. Kantar estimates that 30 to 40 percent of online survey responses are compromised by bots, fraud or professional respondents. Around 3 percent of devices complete 19 percent of all online surveys. Research published in PNAS found LLM-powered bots pass standard survey quality checks 99.8 percent of the time. Faced with panels a third full of noise, a synthetic panel can look like an improvement. It is a substitution of one unverifiable source for another.
When to use which
- Use synthetic augmentation when you have a clean, verified quantitative sample of at least 300 and need to read small subgroups. Disclose it in the report.
- Use LLM respondents to test a questionnaire or generate hypotheses before fieldwork. Never report their output as a finding.
- Use real, verified humans for anything qualitative, anything B2B or expert, anything novel, and anything where the decision is expensive to reverse.
How Theory works
Theory Intelligence uses only verified human sources for evidence. Every interviewee is identity-checked and role-confirmed, every transcript is logged, and the client receives a provenance appendix. We use AI for scheduling, transcription and first-pass coding, and we say so. See how a Decision Study compares to synthetic research, or read The Theory Standard.
Sources
- Fairgen and Google, ESOMAR synthetic data study, February 2026
- User Intuition, The research data quality crisis, February 2026 (citing Kantar, Case4Quality, PNAS)
- GreenBook, GRIT Insights Practice Report 2025