Artificial Intelligence's Geometry Exam: Competencies in Generating Geometry Questions According to Bloom's Taxonomy

This study examines the extent to which generative artificial intelligence tools (ChatGPT, Gemini, and DeepSeek) can generate middle school geometry questions aligned with the cognitive levels of Bloom’s Revised Taxonomy. The study also evaluates the quality and taxonomic appropriateness of these questions based on expert judgments. A convergent parallel mixed-methods design was employed. A total of 30 geometry questions, including open-ended and multiple-choice items, were generated by the AI tools using structured prompts. Ten mathematics education experts evaluated each question using a rubric that assessed alignment with the intended Bloom level. Quantitative data were analyzed using frequencies, percentages, and Gwet’s AC1 agreement coefficient, while qualitative data were analyzed through thematic content analysis to identify patterns in expert justifications. The findings indicate that all three AI tools perform relatively well at lower and middle cognitive levels (understanding, applying, and analyzing), whereas their performance declines at higher levels (evaluating and creating). DeepSeek and Gemini showed higher overall alignment with targeted cognitive levels compared to ChatGPT. However, expert agreement decreased significantly at higher-order levels, indicating inconsistencies in how these levels are interpreted and represented. Qualitative findings revealed issues related to insufficient cognitive depth, weak distractors in multiple-choice items, and mismatches between intended and actual cognitive levels. Overall, the results highlight both the potential and limitations of generative AI in educational assessment and emphasize the necessity of expert supervision in ensuring pedagogical and taxonomic accuracy.This study compares the performance of three generative artificial intelligence tools—ChatGPT, Gemini, and DeepSeek—in generating questions aligned with Bloom’s six cognitive levels in middle school geometry. The participants consisted of ten mathematics teachers with expertise in mathematics education. The teachers evaluated the AI-generated questions for alignment with the intended cognitive level and provided judgments on their taxonomic appropriateness and overall quality. Quantitative analyses examined the agreement between the intended cognitive level and the levels assigned by participants. In addition, qualitative content analysis was used to analyze participants’ evaluations and comments. The findings highlight both the potential and limitations of generative AI tools and emphasize the importance of expert supervision in AI-supported assessment practices.