When AI simulates teaching personas: Linguistic dimensions of LLM-generated classroom discourse
- 1Graduate Program in Applied Linguistics and Language Studies (LAEL) & Graduate Program in Social Psychology (PSO)/Pontifical Catholic University of Sao Paulo, Brazil
- 2Sao Paulo State University at S. José do Rio Preto (UNESP-SJRP), Brazil
Abstract
This study investigates how GPT-4o constructs the social role of teacher through discourse and how this discourse compares with naturally occurring classroom teaching. We generated 19,950 AI classroom teaching texts from a corpus of online college teaching using 56 teacher personas plus a baseline condition. The AI-generated texts contained the original student turns alongside the AI-teacher turns. We analyzed the resulting corpus using Lexical Multi-Dimensional Analysis (LMDA), a bottom-up corpus-linguistic method that identifies dimensions of discourse variation from recurrent patterns of lexical covariation. Five dimensions emerged: Macro-Structuring Discourse vs. Performative Stagecraft, L1-scaffolded Guidance vs. Provocative Challenge, Procedural Control vs. Rapport-driven Engagement, Evaluative Control vs. Reflective Aesthetic Engagement, and Informal Facilitation vs. Formal Deductive Exposition. Comparison with the corresponding human classroom discourse showed that GPT-4o frequently magnified teaching discourses already present in human interaction, particularly organizational, procedural, and expository discourse. It also introduced more artificial forms of teaching discourse characterized by theatrical performance, confrontation, and evaluative pressure. Mixed-effects analyses showed significant differences between human and AI discourse on all five dimensions. Persona prompting also affected dimension scores, but differentiation among the 57 conditions was uneven: many personas occupied overlapping positions, and differences occurred on only some dimensions. The findings show that GPT-4o can produce distinct forms of teaching discourse, yet many different persona prompts repeat similar discourses, suggesting a saturation point beyond which the LLM cannot yield new teaching discourses. More generally, the study demonstrates how corpus-based discourse analysis can characterize the forms of teaching that emerge when an LLM is instructed to perform a professional role.
Related
- A phase field model with stochastic input simulates cellular gradient sensing, morphodynamics, and fidelity of haptotaxis
- Nonverbal Cognition is Associated with Discourse Production in Persons with Aphasia: Evidence from Expository and Picture Description Tasks
- Age-related Differences in Recognition Memory for Discourse: The Case of Modified Words, Competitors, and Related Lures
- Bridging Languages, Rebuilding Communication: Best Practices for Assessing and Treating Discourse for Bilingual People with Aphasia
- Frequency-specific oscillations track parallel processing of linguistic structural and lexical-semantic information from perception to decision
BibTeX
@article{sardinha2026when,
title = {{When AI simulates teaching personas: Linguistic dimensions of LLM-generated classroom discourse}},
author = {Tony Berber Sardinha and Marilisa Shimazumi},
journal = {PLOS One},
year = {2026},
month = nov,
doi = {10.1371/journal.pone.0360650},
url = {https://doi.org/10.1371/journal.pone.0360650},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - When AI simulates teaching personas: Linguistic dimensions of LLM-generated classroom discourse
AU - Tony Berber Sardinha
AU - Marilisa Shimazumi
PY - 2026
DA - 2026/11/06
JO - PLOS One
DO - 10.1371/journal.pone.0360650
UR - https://doi.org/10.1371/journal.pone.0360650
AB - This study investigates how GPT-4o constructs the social role of teacher through discourse and how this discourse compares with naturally occurring classroom teaching. We generated 19,950 AI classroom teaching texts from a corpus of online college teaching using 56 teacher personas plus a baseline condition. The AI-generated texts contained the original student turns alongside the AI-teacher turns. We analyzed the resulting corpus using Lexical Multi-Dimensional Analysis (LMDA), a bottom-up corpus-linguistic method that identifies dimensions of discourse variation from recurrent patterns of lexical covariation. Five dimensions emerged: Macro-Structuring Discourse vs. Performative Stagecraft, L1-scaffolded Guidance vs. Provocative Challenge, Procedural Control vs. Rapport-driven Engagement, Evaluative Control vs. Reflective Aesthetic Engagement, and Informal Facilitation vs. Formal Deductive Exposition. Comparison with the corresponding human classroom discourse showed that GPT-4o frequently magnified teaching discourses already present in human interaction, particularly organizational, procedural, and expository discourse. It also introduced more artificial forms of teaching discourse characterized by theatrical performance, confrontation, and evaluative pressure. Mixed-effects analyses showed significant differences between human and AI discourse on all five dimensions. Persona prompting also affected dimension scores, but differentiation among the 57 conditions was uneven: many personas occupied overlapping positions, and differences occurred on only some dimensions. The findings show that GPT-4o can produce distinct forms of teaching discourse, yet many different persona prompts repeat similar discourses, suggesting a saturation point beyond which the LLM cannot yield new teaching discourses. More generally, the study demonstrates how corpus-based discourse analysis can characterize the forms of teaching that emerge when an LLM is instructed to perform a professional role.
SN - 1932-6203
ER -