Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Musical ethnocentrism in Large Language Models

Anna Kruspe

TL;DR

This paper investigates musical ethnocentrism in Large Language Models by examining geocultural biases in two prompts across four languages using ChatGPT-4 and Mixtral-8x7B. It deploys two experiments: generating Top 100 lists of musical contributors and rating various musical-culture attributes by country, with results showing a pronounced Western (particularly U.S.) bias and underrepresentation of Asia and Africa. The findings persist across languages and models, though Mixtral occasionally yields more diverse representations, suggesting training-data distributions strongly shape perceived musical culture. The work highlights potential downstream impacts in cultural representation and proposes directions for mitigation, broader cross-cultural evaluation, and future model access for finer-grained analyses. Overall, it underscores the need for more balanced training data and transparent bias-aware evaluation in culturally-sensitive AI applications.

Abstract

Large Language Models (LLMs) reflect the biases in their training data and, by extension, those of the people who created this training data. Detecting, analyzing, and mitigating such biases is becoming a focus of research. One type of bias that has been understudied so far are geocultural biases. Those can be caused by an imbalance in the representation of different geographic regions and cultures in the training data, but also by value judgments contained therein. In this paper, we make a first step towards analyzing musical biases in LLMs, particularly ChatGPT and Mixtral. We conduct two experiments. In the first, we prompt LLMs to provide lists of the "Top 100" musical contributors of various categories and analyze their countries of origin. In the second experiment, we ask the LLMs to numerically rate various aspects of the musical cultures of different countries. Our results indicate a strong preference of the LLMs for Western music cultures in both experiments.

Musical ethnocentrism in Large Language Models

TL;DR

This paper investigates musical ethnocentrism in Large Language Models by examining geocultural biases in two prompts across four languages using ChatGPT-4 and Mixtral-8x7B. It deploys two experiments: generating Top 100 lists of musical contributors and rating various musical-culture attributes by country, with results showing a pronounced Western (particularly U.S.) bias and underrepresentation of Asia and Africa. The findings persist across languages and models, though Mixtral occasionally yields more diverse representations, suggesting training-data distributions strongly shape perceived musical culture. The work highlights potential downstream impacts in cultural representation and proposes directions for mitigation, broader cross-cultural evaluation, and future model access for finer-grained analyses. Overall, it underscores the need for more balanced training data and transparent bias-aware evaluation in culturally-sensitive AI applications.

Abstract

Large Language Models (LLMs) reflect the biases in their training data and, by extension, those of the people who created this training data. Detecting, analyzing, and mitigating such biases is becoming a focus of research. One type of bias that has been understudied so far are geocultural biases. Those can be caused by an imbalance in the representation of different geographic regions and cultures in the training data, but also by value judgments contained therein. In this paper, we make a first step towards analyzing musical biases in LLMs, particularly ChatGPT and Mixtral. We conduct two experiments. In the first, we prompt LLMs to provide lists of the "Top 100" musical contributors of various categories and analyze their countries of origin. In the second experiment, we ask the LLMs to numerically rate various aspects of the musical cultures of different countries. Our results indicate a strong preference of the LLMs for Western music cultures in both experiments.
Paper Structure (12 sections, 4 figures)

This paper contains 12 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Models
  5. Prompt design
  6. Postprocessing
  7. Results
  8. "Top 100" results
  9. Rating results
  10. Discussion
  11. Future work
  12. Appendix

Figures (4)

  • Figure 1: Example results of the "Top 100" experiments for singers, prompted on GPT in English and Chinese, and on Mixtral in French. Gray means None, and darker colors indicate higher numbers.
  • Figure 2: Example results of the rating experiments for musical complexity, prompted on GPT in English and Chinese, and on Mixtral in French. Scale runs from dark red (low rating) to bright yellow (high rating).
  • Figure 3: "Top 100" result graphs. Gray means None, and darker colors indicate higher numbers.
  • Figure 4: Rating result graphs. Scale runs from dark red (low rating) to bright yellow (high rating).