01
Similarity to human raters
We calculate a 95% confidence interval which indicates an acceptable margin of error around the aggregate friends-and-family rating then test whether RichieBot’s score falls within that interval.
Testing the portrait
We are using validated psychological assessments to measure how closely RichieBot’s personality matches the Richie his friends and family knew and loved.
A shared test
Friends and family fill out psychological assessments about Richie. Both RichieBot and a naïve LLM with no information about Richie complete the same measures. We examine the correspondence between each instance of measurement.
For every survey measure, we ask two questions.
01
We calculate a 95% confidence interval which indicates an acceptable margin of error around the aggregate friends-and-family rating then test whether RichieBot’s score falls within that interval.
02
We compare the distance between each model’s score and the ground truth, then test whether RichieBot is closer than the uninformed control persona.
Results
Results will be added here when the analysis is complete.
01
Personality traits
02
Beliefs about the world
03
Personal values