Prototype Effects and Implicit Models in LLM Responses to Geometry Tasks

Authors

DOI:

https://doi.org/10.34190/icer.3.1.5686

Keywords:

Artificial Intelligence in Education, Prototype Effect, Implicit Model, Didactical Contract, Geometry Education

Abstract

This paper investigates the role of implicit models in large language model (LLM) responses to geometry questions, with a particular focus on the phenomenon of prototype effects (Rosch, 1975). Building on Fischbein’s account of tacit and implicit models in mathematical thinking (Fischbein, 1987; 1989), together with his later theory of figural concepts (Fischbein, 1993), we analyze how LLMs reproduce and reinforce such models in their explanations. The study draws on a large dataset of elementary geometry tasks in Czech language, as a case of less dominant language. The theoretical framework combines Fischbein’s theory of figural concepts (Fischbein, 1993), emphasizing the dual conceptual–figurative nature of geometric objects, with research on prototype effects in geometry education. These effects manifest, for example, when angles are interpreted primarily as convex regions or when geometric figures are associated with dominant visual prototypes rather than formal definitions (Brunheira and da Ponte, 2019). We consider that LLMs can tend to systematically privilege such prototypical representations. Methodologically, we conduct a qualitative content analysis of LLM responses (ChatGPT and Claude) to geometry prompts. Unlike previous analyses focused on correctness, richness, or errors within Czech and English prompts’ responses, this study codes for the presence, type, and didactical role of implicit models. Particular attention is paid to how these models align with or diverge from local curricular norms and classroom expectations. Results show recurring prototype-based representations, some of which may conflict with local curricular standards despite being mathematically valid. This creates a new layer of didactical tension: not merely between correct and incorrect answers, but between formally valid explanations and those aligned with classroom practice. We discuss this tension in relation to the didactical contract, where AI systems become co-constructors of mathematical meaning. The paper contributes to research on AI in education by showing that LLMs may reproduce and reinforce representations that could influence learners’ implicit conceptual frameworks. This has significant implications for teaching: rather than treating AI outputs as neutral resources, educators must address their underlying implicit models and integrate them into reflective classroom practice.

Downloads

Published

2026-11-13