LLMs Show Sociodemographic Biases in Educational Counselling
A study on arXiv (2604.25932) reveals that Large Language Models exhibit measurable sociodemographic biases when used for educational counselling. Researchers tested six LLMs on 900 student vignettes across 14 identifiers including race, gender, socioeconomic status, and immigrant background, generating 243,000 responses. All models showed biases, partially aligning with human biases but diverging in key ways. Vague descriptions amplified disparities nearly threefold, while concrete metrics reduced bias. The findings underscore risks of deploying LLMs in education without mitigation.
Key facts
- Study examines sociodemographic biases in LLM-based educational counselling
- Six LLMs tested on 900 vignettes with 14 sociodemographic identifiers
- 243,000 model responses analyzed
- All models exhibit measurable biases
- Bias patterns partially align with human biases but diverge
- Vague descriptions amplify disparities nearly threefold
- Concrete, individualized metrics reduce bias
- Published on arXiv with ID 2604.25932
Entities
Institutions
- arXiv