LLMs’ ability to represent value and culture shifts

Determine whether and how Large Language Models can represent value and culture shifts induced by citizen assemblies or cooperative firms, including modeling changes in individual values and broader social transformation, and identify measurable properties of Large Language Models that correspond to such transformations in humans.

Background

The paper discusses the use of LLMs to simulate linguistically sophisticated agents within democratic processes, such as citizen assemblies and democratic firms. It raises concerns about value alignment and the capability of LLMs to capture evolving cultural and individual value changes that such institutions may catalyze.

Because LLMs are trained on text from specific contemporary cultures, their ability to reflect shifts in values and culture over time is queried. The authors explicitly identify as unsolved the problem of representing value and culture transformations and how to measure them in LLM properties linked to human analysis.