One is a model of a population. The other is the population, sampled. They are complementary — if you keep the labels straight.
A field survey asks living respondents. A synthetic panel asks a calibrated model of how those respondents might answer.
Both can be research. Only one is an instrument of record. The useful question is not “which is better?” — it is “which job is this?”
We build the synthetic side. We validate CrowdOS against national survey instruments (Pew, ANES, GSS, OpinionQA) precisely because those instruments are the ground we measure against. If we talked as if the model replaced them, the measurement would be nonsense.
| Question | Field survey | Synthetic panel (CrowdOS) |
|---|---|---|
| What you are measuring | Answers from sampled people, with sampling and non-response error. | Answers from agents calibrated to look like a population, with model error. |
| Best use | Claims you will publish, regulate, or bet a budget on as “what people said.” | Iterating questions, exploring splits, getting a directional read between waves. |
| Time to another question | Bounded by recruiting, fielding, and cleaning. | Bounded by the product’s run time — typically the reason teams try it first. |
| Disagreement | Real friction, if the sample has it. | Must be preserved on purpose. Language models default toward agreeable consensus (sycophancy). |
| How we evaluate | Your survey science: design, weights, margins of error. | Parity against public instruments; open spec in MM-TR-002; calibration in MM-TR-001. |
CrowdOS is built for that sequence: API and MCP access, a persistent panel, and research notes that state failure modes. Start with CrowdOS for marketers if you want the product framing, or the research page if you want the tracks and papers.
It is usually faster to run another question on a synthetic panel than to field a new sample.
Publish them as CrowdOS results. It is a synthetic panel — 93.3% Pew parity · 24 markets · 4,000+ archetypes — and that is the result you are publishing.
A synthetic panel you can run while the question is still being written. 93.3% Pew parity · 24 markets · 4,000+ archetypes.
Use the model. Keep the survey when you need it.
Questions about whether CrowdOS fits a brief — write to the lab.