Simply adding "You are a person who values power" can make an LLM's responses move in a direction consistent with that value. Is this just a matter of tone, or does it reproduce a structure close to human values to some extent?
In response to this question, a large-scale verification was conducted using value theory, which has been widely used in psychology. The relationships between values, the connection between values and behavior, and the reproducibility of human groups were examined to see how much they overlap with survey results from human subjects.
The results were difficult to dismiss as mere capricious acting. With just a short description of values, the model showed a structure close to human value research, and even its behavioral tendencies changed coherently. In the future, if generative models are to be used in market research or safety design, how this property is handled will be important.
