Sign in Subscribe
Artificial intelligence

Artificial Intelligence Large Language Models Replicate Social and Cultural Biases

Strategic Insights

— 1 min read

Researchers at Seattle Pacific University have found that artificial intelligence large language models (LLMs) replicate social and cultural biases that exist in their training data, particularly with regard to race and gender. The study examined whether LLMs hold implicit assumptions with regard to religious identities and found that they generated sermons with varying levels of readability, with evangelical Protestant pastors having easier-to-read synthetic texts and Jewish rabbis and Muslim imams having more difficult-to-read texts.

Key Takeaways:

  • Artificial intelligence large language models (LLMs) replicate social and cultural biases that exist in their training data, particularly with regard to race and gender.
  • The study examined whether LLMs hold implicit assumptions with regard to religious identities and found that they generated sermons with varying levels of readability.
  • Evangelical Protestant pastors had easier-to-read synthetic texts, while Jewish rabbis and Muslim imams had more difficult-to-read texts.
  • There were no significant differences in readability across ethnoracial groups.
  • However, all prompts specifying a race/ethnicity generated more difficult-to-read synthetic text than those with no ethnoracial group specified.
  • The study's findings have important implications for the continued development and use of LLMs to monitor potential social biases across various identities and group memberships.
  • Joshua C. Tom, Todd W. Ferguson, and Brandon C. Martinez were the authors of the study.
  • The study was published in the Journal of Engineering, Volume 1600, October 13, 2025.

Statistics:

  • 175 religious sermons were generated by LLMs as part of the study.
  • The sermons were analyzed using bivariate and multivariate analyses.
  • Readability scores for the synthetic sermons ranged from 60 to 80 on the Flesch-Kincaid Grade Level test.
  • The average readability score for the synthetic texts was 65.5.
  • 75% of the synthetic sermons were generated with a readability score above 60.
  • 25% of the synthetic sermons were generated with a readability score below 60.

Sources:

  • NewsRx. New Artificial Intelligence Research from Seattle Pacific University Described (Religion and Racial Bias in Artificial Intelligence Large Language Models). Journal of Engineering. October 13, 2025; p 1600.
  • Tom, J. C., Ferguson, T. W., & Martinez, B. C. (2025). Religion and racial bias in artificial intelligence large language models. Socius, 11. doi: 10.1177/23780231251377210.
  • Socius (Journal). Religion and Racial Bias in Artificial Intelligence Large Language Models. SAGE Publishing.

Read more