Send message to Authors

Short abstract

Prevalent benchmarks for cultural diversity in LLMs stabilize output resemblance with global value surveys as ground truth. We conduct an experiment to explore whether output-resemblance generalizes to alignment in practice and how users negotiate encounters with nominally misaligned models.

Long abstract

Prevalent benchmarks for cultural diversity in large language models (LLMs) assume that alignment between human and AI can be measured independently of the interaction between them. When a model can match the way culturally diverse humans answer questions on e.g. the World Values Survey, it is taken as an indicator for its capacity for cultural alignment.

However, it is not self-evident that such alignment in output leads to alignment in practice. One could thus reasonably hypothesize both that users interpret and negotiate what counts as value alignment in situ and that output resemblance with value aligned survey responses does not generalize easily to such situations. Thus, current alignment benchmarks of cultural diversity in generative AI likely ignores the sociotechnical contingencies of LLMs in use and prematurely reduces the problem of alignment to something that can be stably measured and easily corrected for.

As part of the Culturally Explainable AI (CXAI) project, funded by the Independent Research Fund Denmark, we develop a gamified experiment where participants are asked to evaluate outputs from models that are deliberately aligned and misaligned with their value positions according to current state-of-the art benchmarks. The aim is two-fold. Firstly, it explores the hypothesis that alignment measured by output-resemblance does not generalize to alignment in practice. Secondly, the experiment functions as prototype for what we envision as future elicitation device to qualitatively investigate the situated sense-making of LLMs in different contexts. We discuss the initial findings and experiences with using this experimental prototype to study GenAI-human relations.