Synthetic Duplicates for a Real Client

Firms are starting to replace survey panels with “synthetic respondents,” language-model personas that answer questions the way a type of customer might. In this course project, students use the technique for a real client and decide for themselves how far it can be trusted. I run versions of it at the University of Tennessee and, earlier, at the University of Kansas.

Illustration: one real person on the left, and on the right a grid of many synthetic copies — Illustration: one real customer, many synthetic duplicates.

How the project works

Illustration: A storefront under a magnifying glass

Step 1

Pick a client and a problem

Teams partner with a local business, student organization, or nonprofit, learn its business, and narrow its needs down to a single marketing problem worth answering.

Illustration: A clipboard survey with rating-scale questions

Step 2

Design and defend a survey

They write the survey that would inform the client's decision, and they have to justify why each question belongs and what it will tell the manager.

Step 3

Build 500 synthetic duplicates

Drawing on conversations with the client and outside research, they build a pool of 500 AI-generated respondents in R, each with its own profile, and have every one of them take the survey.

Illustration: A scatter plot with three clusters, a bar chart, and R code

Step 4

Analyze and recommend

They segment the respondents, model their behavior, and test for differences between groups, then turn the results into recommendations the client can act on.

Illustration: A balance scale weighing a real person against a dashed synthetic one

Step 5

Judge the method

The final presentation and report do two jobs: advise the client, and assess whether synthetic respondents are a viable substitute for real ones, including where they mislead.

Illustration: Three panels: one person, a segment of people, and a bell curve

Step 6

Ask “a duplicate of what?”

Class covers the case for and against synthetic respondents, what replicates and what doesn't, and why a difference between synthetic groups can come from how the personas were built rather than from real customers. Is the duplicate a person, a segment, or a distribution?