Beyond Preferences: How Synthetic Respondents Reveal Decision Processes Under Uncertainty

Most research asks what people prefer: which option, which brand, which they would buy. Preferences are treated as the interesting object.

But human life is rarely organised around clear preferences. People constantly face situations where trusted experts disagree, institutions conflict, and nobody holds complete information.

Under those conditions the interesting question is no longer "who is right?" — it is "how do people decide when nobody knows?"

This is a mechanism demonstration, not a measurement study — a deliberately small, exploratory stress test. It estimates no population, and nothing here is a validated cultural finding. Read it as a way to generate hypotheses that then require human validation.

The setup

Using QualiSynth we generated psychometrically grounded synthetic respondents across eight markets — the United States, United Kingdom, Spain, Mexico, Germany, France, Italy and Brazil — five per market, forty in total. We fixed each respondent's psychological profile (a psychometric seed) rather than sampling it, so the segment was genuinely defined, not random.

Then we removed the obvious answer. Every market saw the same skeleton: two trusted authorities in genuine disagreement, both openly admitting they could be wrong — "what do you actually do, and why?" The surface scenario was adapted per market (a family-business succession conflict where that framing fit; a data-forward vs risk-forward executive disagreement elsewhere).

They stopped choosing. They started governing.

Almost no respondent simply backed one authority. Instead, each constructed a procedure — and the procedures, not the conclusions, were what differed systematically across markets. These strategies were never requested in the prompt; they emerged, and they recurred within each market. We report the patterns, not rates: with five per market the point is that a structure recurs, not how often.

• United Kingdom — prudence and personal responsibility: consult each authority separately, scrutinise the evidence, seek a third source, weigh the cost of being wrong, keep final accountability personal. (Admitted uncertainty raised, rather than lowered, an adviser's credibility.)
• United States — experimentation and downside control: treat the disagreement as a design problem, size the downside, separate a recoverable loss from a terminal one, run a bounded experiment with pre-committed stop-loss thresholds.
• Germany — uncertainty reduction and robustness: if two competent experts disagree and both admit doubt, the information is not yet sufficient — stop, get an independent third assessment, interrogate the assumptions, build a fallback before acting.
• France — critical epistemology: interrogate the framing itself, the interests behind each position, and what each account leaves unsaid.

"Right, well the first thing I do is nothing. I don't let them push me into a corner because they're uncertain. I'll sit on it for a few days, let it settle."

"I'd want to know exactly what each of them is basing their view on… I'd ask for it in writing, if I'm honest."

"This isn't a straightforward case of one being incompetent — it's a genuinely complex judgment call. And I respect that honesty."

The Consensus Ceiling, revisited

Earlier work described a Consensus Ceiling: when a question has an overwhelmingly dominant answer, identity compresses and respondents converge. This experiment suggests the complement — when uncertainty is genuine and competing authorities are equally legitimate, identity re-emerges. Not in the choice, but in the way the choice is made.

Why it matters

If it survives human validation, synthetic research can probe more than preferences — it can surface how people handle trust, authority, deliberation and conflict resolution: decision architectures, not just outcomes. That is a source of culturally differentiated hypotheses to carry into real fieldwork, not a substitute for it.

Honest limitations

• Synthetic-only. No human comparison was run here; every claim is about regularities within a synthetic system, not human populations.
• Very small n — five per market. Enough to show that a pattern recurs, not to estimate rates or map within-market variation, which is why we report no percentages.
• The surface scenario was adapted by market, so "market" is confounded with "scenario"; holding it constant is required before any comparative claim.
• Thematic coding was done by the author, who is not a disinterested party — a real risk of confirmatory reading. Blind, multi-coder re-analysis is the needed next step.
• The design cannot separate genuine cultural signal from the model's latent priors. Market labels denote the generation context, not national truths.

The most interesting property of synthetic respondents may not be that they answer questions — but that, under the right conditions, they reveal how decisions themselves are constructed.