A town with 25 agents
Park and colleagues’ 2023 study placed 25 generative agents in a sandbox environment. Their architecture combined stored experiences, reflection and planning. The evaluation examined the believability of individual behavior and social coordination, including the spread of an invitation to a party.
That is a useful example of an architecture for agent interaction. Believable behavior in this setting does not establish reliable purchase predictions for a business. Read the 2023 paper.
Grounding agents in participant information
A separate study by Park and colleagues built agents from interviews, surveys or both involving 1,052 Americans. The evaluation included held-out survey responses, with participants’ own two-week answer consistency used as a benchmark.
The revised paper distinguishes the results of the different input methods. These conditions are important when interpreting its findings. Read the revised paper.
What a result does not establish
A generated demographic sketch differs from the interviews and survey responses of a real participant. A result measured with one input and evaluation setup cannot be assumed to apply to another.
Neither study tested Founder Effect. The reported results are not a general accuracy score for synthetic customers, and this site does not use them to promise a percentage chance of commercial success.
How we use the research
Founder Effect is developing a method that combines source evidence, synthetic persona reactions and a separate challenge to the findings. A generated reaction may suggest an objection worth investigating. It still needs an appropriate check.
The output is a set of hypotheses and possible tests, with the evidence and limitations attached. The software platform is not publicly available.
Read the method and research limitations.