What Is a Silicon Sample?
A silicon sample means feeding a large language model a target population's demographic and psychographic makeup, then treating its outputs as a stand-in for how that population would answer research questions. Where a traditional sample means recruiting and surveying 500 real people, a silicon sample generates and queries 500 language-model outputs instead. The economics flip: minutes instead of weeks, a running query instead of a per-study field budget.
Where the method comes from
In 2023, Argyle, Busby, Fulda, Gubler, Rytting, and Wingate published the paper that grounds the field, titled Out of One, Many, subtitled Using Language Models to Simulate Human Samples (Political Analysis, Cambridge University Press) (Cambridge University Press). Their setup: take a frontier LLM, give it the demographic backstory of a real ANES respondent, where ANES is a benchmark survey of US political attitudes, and have the model answer as that respondent would, then aggregate the results across many conditioned samples. Their finding: the resulting opinion distributions tracked the real ANES distributions closely on consistent attitude clusters, such as party affiliation, ideology, and policy preference, and less closely elsewhere. Political science, sociology, marketing, and economics picked up the citation, and that follow-on work turned silicon sampling into a named, studied method.
How a research-grade sample gets built
Five steps recur across implementations:
- Define the target population: the demographic and psychographic parameters that matter, such as geography, age, income, occupation, attitudes, and prior brand exposure.
- Determine sample composition: build proportions along those parameters so the sample mirrors the real population's makeup, rather than generating a flat batch of respondents.
- Calibrate against prior real data: condition on panel data, prior survey waves, or CRM segments where available. This step separates a research-grade sample from a thin model wrapper.
- Generate the sample: produce the conditioned, addressable units.
- Query the sample: submit the research instrument, aggregate, analyze.
What it answers well, and what it doesn't
| A silicon sample is strong for | A silicon sample is not built for |
|---|---|
| Directional reads on opinion and preference, such as ranking concepts, gauging message resonance, or reading brand attitude | Statistically validated population estimates with a defensible confidence interval |
| Hard-to-reach audiences: senior B2B buyers, regulated professionals, future customer segments | Genuinely novel categories with no analog in the model's training data, where output looks plausible but carries no real signal |
| Multi-market comparison run in a single sitting instead of spread across months | Sensory and emotional response: a model can reason about packaging or a TV ad, but it cannot perceive it |
| Continuous, low-cost iteration on the same research question | Any claim of statistical accuracy without stating the method's error and calibration limits |
The right-hand column is the gap that matters for a launch, pricing, or messaging decision. A directional read tells a team what a population is likely to prefer. It does not establish that the preference is real, how large the effect is, or how confident the team should be before committing spend.
A typical deployment pattern
One historical example used a 200-persona silicon sample to screen a dozen concepts and narrow the field to 2 to 3 candidates, compared the same campaign across 4 to 8 country silicon samples in a single sitting before committing media spend, and, before launch, confirmed the surviving 1 to 3 options with a small study of real respondents. Treat these as inherited planning figures illustrating how the method has been sequenced elsewhere, not a current Subconscious product specification, benchmark, or guarantee.
Closing the gap: from directional signal to a decision you can defend
Subconscious is a causal behavioral platform: controlled discrete-choice experiments run against a simulated population, returning causal effects with confidence intervals rather than a single directional read. That answers a different question than "what did the sample say": it answers "what caused the preference, and how much would changing the input change the outcome."
Where a decision genuinely depends on it, a team can move from a simulated experiment to real-human validation without changing the causal question being tested. Subconscious can validate studies with real human participants or, where audience scale matters, run controlled studies against a person-level audience graph covering 800 million real people. That audience graph is a basis for structuring a study at scale, not a recruitable panel of people waiting to be surveyed.
Terminology varies, the method stays the same
Commercial platforms describe this underlying method with several different labels. Some name the group, others name the individual unit inside it. The framing changes depending on whether the source is academic literature, a platform's marketing page, or a sales deck. The mechanism underneath stays constant: condition a model on a target population, generate outputs, aggregate.
What this means for a pending research decision
Before treating a silicon-sample readout as a green light for a launch, price, or message decision:
- Confirm it was calibrated against real prior data from the target audience, not a flat, uncalibrated batch.
- Treat the output as directional signal, not a population estimate. It carries no confidence interval unless the platform generating it produces one.
- Where the decision is expensive to get wrong, route it through a causal experiment, and add real-human validation before committing budget.
Related reading: research, how a study moves through this process, the leaderboard, about Subconscious.