Physical Review Physics Education Research

Using Large Language Models for automated grading of handwritten physics exams is extremely promising, but some problems still remain.

ChatGPT can approach human-level accuracy in grading physics problems, and can be enhanced with multiple trials and explaining the rubric.

A comparison of different LLMs in coding lab notebooks for the identification of presence of skills.

AI can reduce the grading burden for instructors by accurately handling routine tasks, but human oversight remains essential.

Sixth grade students in a randomized controlled study showed some affective benefits when learning with supplementary material generated by AI chatbots but limited learning performance.

Both human-human and human-AI collaboration significantly improved problem solving, but human-human collaboration showed a larger effect.

AI tools can be used to identify and distinguish student trajectories as they learn about energy.

Even when providing correct solutions to an introductory-level energy problem, LLMs fail can fail to generate the assumptions necessary for the solution.

Qualitative analysis shows that ChatGPT-4o struggled on items where representational or visual interpretation is important to generating correct answers.

Physics instructors’ adoption of generative AI is influenced by their level of access and motivation.

Effectiveness of high school students’ use of ChatGPT during a physics laboratory is significantly influenced prior physics knowledge and teacher guidance plays a crucial role.

Extending prior work, the authors find that student’s gaze during conceptual tests more accurately predicts performance than during a learning phase.

Teachers use AI differently in lesson planning according to their needs and experience teaching physics.