Abstract
Generative AI is now a routine study partner in programming courses, yet the emotional tone students use when prompting these systems remains underexamined. This study analyzes student→AI prompts from four academic terms spanning SQL and Python courses. We extract prompt-only text from DOCX/PDF submissions, de-identify it, and classify tone with a fine-tuned DistilBERT model into six categories: Neutral, Polite, Demanding, Frustrated, Tentative, and Curious. We then aggregate tone distributions by student, course, and term. Results show a shift from Neutral toward Curious over time, persistent prevalence of Demanding prompts, and a late-term rise in Frustrated prompts, consistent with “productive struggle” on harder tasks. We discuss implications for tone-aware tutoring (e.g., scaffolding when frustration spikes) and course practices that encourage inquiry-oriented, reflective prompting.
Recommended Citation
Gulzar, Talha and Hashmi, Nada, "Signals in the Prompt: Classifying Tone in Student–LLM Dialogues" (2025). NEAIS 2025 Proceedings. 4.
https://aisel.aisnet.org/neais2025/4