Generative artificial intelligence (GenAI) is transforming physics education by enabling automated solutions and personalized feedback. Project Ethel at ETH Zurich explores integrating GenAI into physics instruction and assessment, leveraging both reasoning and non-reasoning multimodal models.

Our studies show that GenAI can not only pass introductory physics courses but also consistently outperform the average undergraduate on post-instruction physics concept inventories. Models still struggle with tasks requiring visual interpretation. In instructional use, these limitations are particularly pronounced when providing feedback on homework and grading exams, where accurately recognizing handwriting and interpreting graphical diagrams remain challenging.

To address this, we applied psychometric methods - most notably item response theory - to derive confidence measures for GenAI-generated grades. This approach achieved high reliability for subsets of tasks, significantly reducing the need for manual grading.

These findings highlight GenAI’s potential to enhance physics education while underscoring the need for human oversight, curricular adaptation, and attention to equity. The next-generation Ethel platform integrates these capabilities and is available as open-source software to foster cross-institutional collaboration.