Say vs do

You are deciding whether to launch the new pack, and you will greenlight it only if most shoppers would switch to it. So you need to predict how many will. You can ask people in a survey, or ask an AI model that stands in for them. Both can only report what someone says they will do, so watch what happens when you finally check those answers against what shoppers actually did.

The launch rule: greenlight only if more than half of shoppers would switch.

Survey: people say they'll switch64%
Model: it says they'll switch68%

Answered in one second, from text alone. It never saw a shopper.

What shoppers actually did22%

Dashed line is the 50% launch rule. Above it, you greenlight.

94% match. The model reproduces the survey almost exactly. This is how synthetic-respondent tools are usually sold as validated: they line up with human answers.
But look at what that check actually was. You compared one stated answer to another. Say against say. Neither has touched a behaviour yet.

Illustrative example, not recorded model output. The numbers are patterned on the real magnitudes in the article: the human say-do gap (Sheeran, 2002; Murphy et al., 2005), the machine divergence where a fluent stated answer sits above behaviour (Goli and Singh, 2024; Brand, Israeli and Ngwe, 2024), and the vendor pattern of validating synthetic output against human survey answers rather than behaviour. Built 14 August 2026.