Comparing AI Interaction Logs and Classroom Analytics for Detecting Conceptual Difficulties in Statistics

Generative AI tools are increasingly used by students of statistics, but it is unclear whether the questions students ask them reveal the conceptual difficulties that classroom assessment already detects. Classroom polling surfaces a misconception only when a question happens to probe it, so a source available throughout a course could in principle complement it. This exploratory case study asks which topics generate the most AI interactions, which show the highest classroom-response error rates, and how far the two sources agree. We
compared 728 AI-log items (162 conversations, 58 students) from a curriculum-bounded platform with 18 Wooclap items from three tutorials in a second-year statistics course at a Dutch university, with final-exam results as a third reference point, using a shared six-topic coding scheme. AI-log volume did not align with Wooclap error rates: among the three topics present in both sources, Type I/II errors and statistical power had the fewest AI conversations (4) but the highest error rate (57.4%), and point estimation, with the highest exam error rate among comparable items, had no Wooclap item. The comparison is descriptive and limited to one course. It suggests that AI-log volume shows where attention concentrates rather than where students struggle, and that the two sources are best read together.