A Comparative Computational Study of Bias and Affective Modelling in LLM-Generated Feedback on Academic Writing

Márquez Cruz, Manuel

doi:10.22055/rals.2026.52654.3878

A Comparative Computational Study of Bias and Affective Modelling in LLM-Generated Feedback on Academic Writing

Document Type : Research Article

Author

Manuel Márquez Cruz

Department of Linguistics, Arab, Hebrew, Basque and East Asian Studies, Faculty of Philology, Complutense University of Madrid, Madrid, Spain

10.22055/rals.2026.52654.3878

Abstract

This article examines gender bias in feedback generated by three Large Language Models—ChatGPT, Copilot, and Gemini—on forty essays written by second-year undergraduate students enrolled in English for Business Management I at the Complutense University of Madrid. Using an exploratory comparative design that combines qualitative discourse analysis with complementary quantitative comparisons, the study analyses not only marks but also the discursive and emotional construction of evaluative judgement, focusing on tone, mitigation of criticism, positive reinforcement, process orientation, and the implicit image of the student projected by each model. The findings suggest that, within this corpus, potential gender-related differences are manifested less in numerical scores than in the evaluative voice of the feedback. Feedback addressed to female students tended to be more mitigated, supportive, and process-oriented, whereas feedback addressed to male students tended to be more direct, technical, and corrective. Within this corpus, the three models appear to articulate these gender-related patterns in different ways.

Keywords

Artificial Intelligence

Educational Feedback

Gender Bias

Large Language Models

Higher Education.