Would people trust a decision made from AI-simulated opinions?
Synthetic audiences let teams pre-test messaging, pricing and concepts without fielding a real survey. The obvious question: do people actually trust decisions made that way? We put it to a synthetic audience grounded in US population structure — fittingly, a synthetic read on trust in synthetic reads. Directional, not a real poll.
“Would you trust a company's product or pricing decision that was based on AI-simulated customer opinions instead of a real survey?”
Margin of error on the topline ±4 points (95% confidence, 400 responses).
- 78% chose "Only as an early signal, then verify with real people" (±4 points, 95% confidence).
- Second most common: "No — real people only" at 20%.
- By age group: "Only as an early signal, then verify with real people" was strongest among 35-44-year-olds (84%) and weakest among 55-64-year-olds (71%).
We put the question "Would you trust a company's product or pricing decision that was based on AI-simulated customer opinions instead of a real survey?" to a synthetic audience of 400 people, grounded in US microdata (US Census ACS 1-year PUMS microdata).
The audience represents US adults.
The most common answer was "Only as an early signal, then verify with real people" at 78% (±4 points).
Full distribution: Only as an early signal, then verify with real people (78%), No — real people only (20%) and Not sure (2%).
By age group: "Only as an early signal, then verify with real people" was strongest among 35-44-year-olds (84%) and weakest among 55-64-year-olds (71%).
By education: "Only as an early signal, then verify with real people" was strongest among the "Bachelor degree" group (86%) and weakest among the "High school" group (69%).
By income: "Only as an early signal, then verify with real people" was strongest among the "$150k+" group (96%) and weakest among the "<$25k" group (71%).
The honest positioning matches what most people intuitively want: use synthetic responses as a fast, cheap first read to pre-test and narrow options — then validate the high-stakes calls with real people.
The tools that will earn trust are the ones that show their own accuracy. A synthetic audience built on real census microdata, with an explicit calibration score, tells you exactly how far to trust each result.
This is a synthetic study, not a real survey. Answers come from an AI audience grounded in US real microdata (age, sex, education and income jointly) (US Census ACS 1-year PUMS microdata). The distribution has genuine spread (2 options with meaningful support), but still treat it as a quick directional read, not a measured fact.
For high-stakes calls, use this as a first direction and validate with real people before you commit.
This is a synthetic study, not a real survey. Answers come from an AI audience grounded in US real microdata (age, sex, education and income jointly) (US Census ACS 1-year PUMS microdata). The distribution has genuine spread (2 options with meaningful support), but still treat it as a quick directional read, not a measured fact.
Are synthetic respondents accurate?
They're directional, and accuracy varies by topic. The useful ones are grounded in real population data and report a calibration score so you know when to trust them and when to validate with a real study.
When should I use synthetic vs a real survey?
Use synthetic to pre-test fast and cheaply — messaging, concepts, pricing, naming. Validate high-stakes decisions with real fielding. They're complements, not replacements.
Ask any question to a representative synthetic audience and get a simulated response distribution — grounded in real census microdata.
Method: 400 synthetic respondents, gpt-4o-mini, grounded in US microdata (US Census ACS 1-year PUMS microdata). Published October 2026. Generated with Chorus.