Are AI synthetic audiences accurate?
Yes, for seeing reactions coming before an announcement. No, as a stand-in for a poll. A synthetic audience is a group of characters created by an AI who react in place of real customers or employees.
Don't ask it for a percentage. Ask it who will push back.
A tool that tells you "this share of your customers will accept the price increase" is selling you false precision: against a real poll, the gap runs from 4 to more than 23 points depending on the topic. But the day before an announcement, knowing which group will take it badly, over which sentence, and what to fix, is exactly what simulated voices do well.
A serious tool shows its results
One question sorts the market: "Show me your results on real cases the tool had never seen." And be wary of anyone promising to predict the future.
100 real decisions, 77% of the criticism found
of the criticism raised at the time, found on average by Kapari without knowing how the story ended.
real past decisions, 50 in France and 50 in the US, put through the exam on October 21, 2026.
In practice: before your announcement, you know most of the objections you are going to hear, and you have time to answer them. Kapari will not tell you how many customers you will lose; it tells you who, and why.
For the curious: the sources
- Verasight reports, 2025-2026: a gap of 4 to more than 23 points between a simulated group and a real poll.
- Bisbee et al., Political Analysis, 2024: on a simulated group whose averages looked right, 48% of the links between answers differ from reality.
- Santurkar et al., ICML 2023; Wang, Nature Machine Intelligence, 2025: AI erases minority views.
- Kapari's exam, case by case: the exam page.
- To place these tools in the market: our guide to synthetic audiences.