ChatGPT-4’s ability to describe kinematic graphs

Sammanfattning

This study investigates whether ChatGPT-4V can generate textual descriptions of kinematics graphs that could be used by students with visual impairments. The model produced clear and accurate descriptions for some simpler graphs but struggled with more complex ones. Recurring problems included incorrectly stating that graphs began at the origin, misidentifying or omitting intersections with the x-axis, and occasionally introducing answer-oriented hints despite explicit instructions not to do so. In addition, the quality of the descriptions did not consistently correspond to the model’s ability to correctly answer the same items in prior work. These findings suggest that ChatGPT-4V is not yet reliable as a stand-alone accessibility tool for describing kinematics graphs. However, the model may still have potential as a support tool when used with expert review and further improvements to multimodal AI systems.