When Synthetic Respondents Stop Choosing

Most of the debate about synthetic respondents is stuck on one question: do they sound like real people? It is a fair question. It is also the least interesting one, because the answer is mostly "yes, fluently" — and fluency is exactly what should make you suspicious.

So we asked something else.

Not which option would you choose? — that question has an obvious answer in most markets, and obvious answers make everyone converge. We asked the harder version:

What would you do when two people you trust disagree, and neither of them is certain?

That small change turned out to matter more than any model upgrade. When the answer stopped being obvious, the respondents stopped choosing — and started constructing a process. And the process was different in every market.

The setup

Eight markets — UK, US, Spain, Mexico, Germany, France, Italy, Brazil. Five synthetic respondents each, generated under local market norms. One open-ended dilemma per market, always the same underneath: two trusted authorities in genuine disagreement, both admitting they could be wrong. No correct answer baked in.

This was a demo of the engine, not a validated study — we will be honest about that throughout, because it is the whole point.

What we found

Almost nobody just picked a side. Nearly everyone built a procedure. But the procedures diverged in a way that was hard to unsee:

The British deliberated. The Americans engineered an experiment with a stop-loss. The Germans hit the brakes and asked for more certainty. The French interrogated the assumptions behind each expert rather than the experts themselves. The Mexicans turned the conflict into a conversation designed to protect relationships. The Brazilians did the same — but to learn together, not to keep the peace. The Italians refused the choice entirely and asked what can change without emptying the business of meaning. The Spanish piloted everything: "escucho a los dos, pero mando yo."

Same question. Eight different philosophies of what a good decision even is.

Why this matters for research

Better synthetic research depends less on a better model and more on a better dilemma.

When you ask a question with an obvious answer, synthetic populations collapse onto the default. We call that the Consensus Ceiling. But when you build genuine tension between two legitimate positions, they have to construct a path — and that is where the differentiated, usable material lives. The output stops being a preference and becomes a decision architecture: who they trust, what evidence they want, whether they pilot or postpone, whether they protect relationships or accept risk.

For pre-research and concept work, that is the more valuable signal. Preferences are outcomes. Architectures are how the outcome gets made.

The honest part

We have to be clear about what this is and is not, because the result is seductive and easy to over-read.

These are synthetic respondents only. No humans were compared in this study. Five per market is a demonstration of a recurring pattern, not a population estimate. And the most important caveat: the responses are culturally legible — the Spanish "pilot first," the German "stop and reduce uncertainty" — which is exactly what you would expect if the model were rendering cultural stereotypes rather than surfacing real structure. On this design, we cannot fully separate the two.

That is not a footnote. It is the reason we treat every architecture above as a hypothesis to take to human fieldwork, never as a description of how a country decides.

Try it yourself

This whole study came out of the public playground — a plain-text segment description and one good question. You can run the same kind of dilemma on your own segment in a few minutes.

A full working paper — method, related work, and the full treatment of the stereotype problem — is in preparation. Powered by the StrataSynth engine.