The Assessment of ChatGPT and DeepSeek to Solve Chemistry Exams and the Classification of the Exam Questions According to Bloom’s Taxonomy in order to Alter the Level of Exams. A Comparative Study

Abstract

This study examines the performance of ChatGPT and DeepSeek in solving undergraduate chemistry exams (general, organic, inorganic, and physical) and their ability to classify questions by Bloom’s Taxonomy to adjust cognitive levels. Both models perform well on general chemistry, with DeepSeek scoring about eight points higher, and both succeed on direct calculation tasks (e.g., kinetics, atomic structure, and Bohr model). ChatGPT struggles with complex numerical problems (e.g., Arrhenius equation) and does not handle diagram based tasks (e.g., orbital diagrams). In physical chemistry, both achieve near-perfect scores on thermodynamics questions, while performance on organic chemistry is low (38% for ChatGPT vs. 23% for DeepSeek), reflecting difficulty with stereochemistry, acid-base chemistry, resonance, and spatial structures. In inorganic chemistry, both perform moderately (62% vs. 56%), solving simple but not symmetry or group theory-related problems. Statistical analyses across multiple-choice questions show that differences between models are small and not statistically significant for any exam: organic chemistry (37.5% vs. 25.0%, p = 0.63), inorganic (84% vs. 68%, p = 0.72), general (100% vs. 86.7%, p = 0.50), and physical chemistry (100% vs. 95.2%, p = 1.00), indicating only potential advantages that cannot be generalized. Both classify most questions correctly by Bloom’s levels, though errors occur at the apply/analyze levels. ChatGPT is more effective in adjusting question difficulty by moving through Bloom’s verbs.