The recent ability of ChatGPT-4 to process image data as input led us to test its performance in interpreting graphs, one of the most common representations used in physics. We tested ChatGPT’s performance on the well-established Test of Understanding Graphs in Kinematics (TUG-K). The study shows that ChatGPT's responses are quantitatively similar to those of students. However, an in-depth analysis reveals how they qualitatively differ. The chatbot often succeeded in devising strategies for the tasks and accurate “reasoning,” but failed in correctly “seeing” graphs. On the contrary, students typically do not have problems interpreting graphs but struggle with correctly reasoning. The findings suggest that ChatGPT’s abilities are not human-like, and this represents a challenge for its use in physics education.