Silicon Sampling

Origin

Argyle and colleagues coined the term in 2023 in the journal Political Analysis. They gave a language model thousands of sociodemographic backstories of real participants from US surveys and had it answer their questions. For many subgroups, the distributions of the simulated answers were close to those of the real survey. The authors called this property algorithmic fidelity.

The study was deliberately designed as a proof of concept: can a model serve as a proxy for certain population groups, for example to pre-test questionnaires or form hypotheses?

Why this matters for your decision

When a provider presents synthetic respondents to you, there is usually a variant of this principle behind it. For your assessment, what matters is what the agreement was measured against. In silicon sampling, the yardstick is a survey: the model matches what people state in surveys. Whether it matches what people do when a new offer with a price is in front of them has not been tested.

A language model reproduces what people have said and written. For familiar topics with a large body of text, this can work surprisingly well. For offers that do not yet exist, this basis is missing.

In practice, this means: silicon sampling is a good candidate if you would otherwise have set up a quick survey on a familiar topic, for example to test wordings or segments. It is a weak candidate if the question is how many people will choose a new offer at a specific price.

Example

A telecommunications provider wants to know how households in three regions react to a fibre tariff with a rented router. A silicon sample of 1,000 synthetic profiles delivers within a few minutes: 38% interest overall, higher among younger homeowners. That is useful for sharpening segments and questions for the next step.

Whether the same households would sign up for the tariff if they saw it in their familiar online environment next to their current connection is not answered by this figure. That requires measured sign-up intent from real people.

Distinction

Silicon sampling is a method within synthetic research. Synthetic respondents, AI personas and synthetic consumers describe the result, that is, the simulated people. A digital twin usually replicates a specific target group or customer base with its own data. Synthetic data in the broader sense also includes artificially generated tabular data with no connection to surveys.

Limits

Follow-up studies show where the method reaches its limits. In 2024, Bisbee and colleagues found matching means, but considerably less variance than in the real survey, deviating relationships between attributes and noticeable fluctuations of the same prompt over three months. In addition, model responses most closely resemble US and Western patterns, which is relevant for markets in Europe.

Horizon therefore classifies silicon sampling as a tool for exploration: for hypotheses, segments and an initial pre-selection. The final variants go into a behavioural test with real people before the investment.