What works best
How accurate a file is depends mostly on what you ask, not on the answer scale. In our tests, the same answer format could be very accurate for one topic and rough for another. We checked this directly: when the same questions were asked with 2, 6 and 10-point scales, the AI's errors stayed with the questions, not the scale.
Most reliable
- Opinions on broad social and political issues (equality, government, trust in institutions).
- Personality and self-description ("I am the life of the party", "I get chores done right away").
- Factual behaviors and circumstances asked as yes/no or select-all (have you done X; which of these do you use; banking and payment habits).
- Work facts asked as numbers (hours worked, commute time): the typical answers and the spread come out close to Census figures.
Use with more caution
- Mental health and wellbeing (how often people feel anxious, down, stressed).
- Social support (how often someone is there for you).
- Past experiences (what people worried about as children, things that happened years ago).
- How much people value specific things (importance of nearby amenities, for example).
- Narrow groups and politically charged questions.
The common thread: the AI does well with opinions and with facts about behavior, and less well with people's inner experience and personal circumstances, where its picture of people differs from what people report. For these topics, check key results with real respondents before relying on them.
Answer formats we offer
We only offer answer scales we have tested: agree–disagree (5 and 7 points), "describes me" (5 points), satisfaction (5), importance (4), how-much-bothered (4), yes/no, never–always (5, our roughest), select-all-that-apply, and number answers. Unordered single choices (e.g. "what kind of car do you drive") and open text are not supported yet.
How sure we are
This guidance comes from comparing synthetic files with held-out real surveys of US adults (election and civic surveys, a Federal Reserve household survey, a national health information survey, Census records) and with large personality, interest and mental-health datasets. Several topics rest on one study and a handful of questions, so treat it as a guide, not a guarantee. We will update it as we test more.