Building Smart Classrooms Through AI and LLMs: A Case Study Using PISA Data
Large Language Models (LLMs) such as ChatGPT and Gemini are reshaping education through personalized learning and data-driven decision-making. This study evaluated LLM classroom feasibility by linking them with PISA student-level data. Using R-based processing, 1,000 structured profiles were submitted to ChatGPT and Gemini APIs to identify academic and socio-emotional strengths, risks, and improvement strategies. Unlike systems that rely on predictive modeling, this approach uses LLMs as interpretive assistants that generate context-aware feedback from standardized educational data without custom model training. Thematic analysis revealed strong cross-model agreement $(\mathbf{r}=\mathbf{0. 8 7})$ and alignment with teacher evaluations ($\mathbf{r}$ = 0.76). ChatGPT provided concise, action-oriented feedback; Gemini offered richer contextual explanations. High inter-rater reliability $\kappa=0.79-0.82)$ confirmed consistent content interpretation. However, the reliance on agreement-based validation introduces a circular reasoning risk that limits causal claims. Findings demonstrate LLMs' potential as real-time educational assistants for early detection and personalized guidance, though input data dependence, bias risks, and limited longitudinal validation remain key limitations.