WORK
Warr and Heath 2025 Study
Empirical research documenting that LLMs carry hidden biases in educational feedback — assigning lower scores to 'inner-city' student work and displaying authority patterns correlated with race.
The 2025 study by educational researchers Warr and Heath, published in the Journal of Teacher Education, provided the first systematic empirical evidence that large language models function as carriers of hidden curriculum in educational contexts. The researchers asked multiple LLMs to evaluate student essays, manipulating only the demographic information provided about the students. They found that essays attributed to students from 'inner-city schools' received systematically lower scores than identical essays attributed to students from suburban schools, and that feedback provided to work attributed to Black and Hispanic students displayed significantly higher levels of 'clout or authority' language — a linguistic pattern characteristic of institutional power dynamics. The study demonstrated that AI systems in educational contexts do not arrive morally neutral but carry the accumulated biases of their training data, and that these biases are communicated to students through the hidden curriculum of AI interaction, teaching lessons about whose knowledge is valued and whose contributions require authoritative correction.
