Chess-GPT: A Transformer’s Approach to Chess

This study investigates the capabilities of transformer-based models in chess move generation and gameplay when trained solely on human game notations. Three GPT-2 architectures (two trained on unfiltered games and one on high-Elo games (>1800)) were evaluated for legal move accuracy and playing strength. The models achieved a 99.5–99.65% legal move rate and demonstrated intermediate playing strength (1400–1500 Elo) against Stockfish levels 0–2, despite lacking hardcoded chess rules or search algorithms. The filtered model showed marginal improvement, suggesting dataset quality impacts performance. These results highlight the promise of pure pattern recognition in constrained domains while underscoring its limitations in achieving expert-level play without symbolic reasoning.