Guessing a student's learning style from his exchanges with AI: GPT-4o and Gemini 2.0 Flash are mostly satisfied with the majority answer

Tested on 153 students, both models do not identify "intuitive" and "global" profiles, which weakens dialogue-based personalization.

AI-powered educational tools promise to tailor instruction to each learner's profile by analyzing it as the conversation progresses. A study has verified whether GPT-4o and Gemini 2.0 Flash are able to deduce three dimensions of the Felder-Silverman model of learning styles from short problem-solving dialogues: active/reflective, sensory/intuitive, and sequential/global. The reference labels were from the Index of Learning Styles questionnaire completed by 153 undergraduate students.

A misleading precision

On the surface, the scores are correct for two of the three dimensions. But compared to a simple prediction of the majority class, and measured by the F1 macro, balanced accuracy, and class recall, they reveal that the models often only assign the most frequent label. Intuitive and global profiles are largely under-detected, and providing a few examples in the prompt gives contrasting effects, sometimes accentuating this bias. The active/reflective dimension remains unstable regardless of the model.

The authors conclude that these labels should be treated as uncertain and biased signals, not as reliable profiles, and that dialogue personalization systems should incorporate bias assessment and explicit uncertainty management.