NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Occupational Prompting Reveals Cultural Bias in Large Language Models
TL;DR
Cultural biases in large language models (LLMs) are not fully understood, particularly how occupational identities influence their responses. This study employed occupational prompting, using professional roles like accountant and teacher, to evaluate LLM responses to value-survey questions.
✦ Why It Matters
Engineers can leverage these insights to mitigate cultural biases in AI systems by considering occupational contexts.
Key Takeaways
- Explore how occupational identities affect AI responses.
- Utilize survey-grounded evaluation methods for bias assessment.
- Consider the implications of cultural biases in AI design.
- Develop strategies to address biases based on professional roles.
Full Summary
Cultural biases in large language models (LLMs) can shape their responses to social and ethical questions. This research introduced a method called occupational prompting, which involves using specific professional identities to assess how LLMs respond to value-based survey questions.
The study utilized a survey-grounded evaluation pipeline based on the Integrated Values Surveys and mapped responses onto the Inglehart-Welzel cultural space. Findings revealed that LLMs, when prompted with occupations, exhibited a consistent Western bias, but responses varied significantly depending on the specific occupation.
For instance, responses from an engineer differed from those of a nurse, indicating that occupational identities elicit structured value patterns. This work expands the understanding of cultural bias in LLMs beyond nationality and provides a framework for further exploration of how professional roles influence AI behavior.