When not use AI generated images in physics teaching

When not use AI generated images in physics teaching Jiří Šperka Mendel University in Brno, Czech Republic Introduction The use of generative artificial intelligence (AI) based on large language model in physics teaching and education has been already studied from various perspectives. E.g. Gerd Kortemeyer1 explored how ChatGPT deals with representative assessment content of an actual calculus-based physics course and grading. W. Yeadon and T. Hardy2 assessed its ability to answer 1337 physics exam questions using gpt-3.5-turbo model suggesting recommendations for educators in addressing AI, concluding that limitations persist in AI application to physics. In.a.recent article,3 it was discussed how generative AI can work in physics classroom. In the area of AI generated pictures with physics content, latter article3 investigated generated images created by Almanack tool, finding pictures on.the slideshow bizarre, also describing free-body diagram generated by ChatGPT-4 full of.nonsensical labels, such as "Gartisal force." M..G. de Souza, et al.4 investigated how school students utilized AI generated images using AI chatbot to create visual representations of."relativity" to represent their understanding of general relativity concepts. Following text is focused solely on the content of AI generated images created using AI chatbot for the purpose of their use in physics teaching, i.e. use of AI generated images that cover some physical phenomena. Text part of this article is written without any help of AI, the AI generated images are the only AI generated component here. The AI pictures were generated using Microsoft Copilot (2024) with conversation style option set to more precise mode (i.e. most precise from three available options, other being more creative and more balanced). Teacher would preferably use AI tool that is free of charge, one of such is freemium chatbot Copilot employing also text-to-image model. One of basic concepts in physics education are mass, length and time. They can serve well as a simple model examples of physical phenomena that one could be tempted to approach with the help of AI. Let's qualitatively investigate how AI deals with those concepts in the area of image generation. AI generated example physics images Three prompts aimed at generating of images corresponding to concepts of mass, length and time follow. First prompt input for Copilot was: Generate physically accurate image in which small cat is weighed against the weight of 50 kg on mechanical weighing scale. I expected generated image in form of two pan balance with small cat on one pan and weight of 50 kg on another pan of the balance. The result is in Fig. 1. There are four images in this figure because Copilot generates four images as the answer on one single prompt. There are more things that are not satisfactory in.generated images: Any of images doesn’t include any weight of 50 kg. Cat isn’t weighted against any weight, cat is simply weighted on bench weighing scale instead. There are two cats in two pictures and we can see clearly weighing of two cats together in one picture, although the aim was to visualize weighing of one cat only. Other errors in images are: Bad dial face with non logical sequence of numbers, unreadable numbers, wrong physical unit KG and G, someone touching weighing scale during the weighing of the cat. Second prompt input was: Generate physically accurate image of clock measuring time with time scale. Response to this second prompt is in Fig. 2. Again, there are errors in images, such as repeatedly bad dial face with non logical sequence of numbers, e.g. sequence XII, I, III, III of.Roman numerals in first image, and unreadable numbers. Generated clock are maybe more valuable from artistic viewpoint rather than chronometry. Figure 1: AI generated images for first prompt. Third prompt input was: Generate physically accurate image of measuring length of 8 cm long pencil using a ruler. We can see generated result in Fig. 3. There are some clear AI hallucinations in.generated pictures: The pencil doesn’t look like real pencil in two cases because of weird tip. In.every picture, the markings on the ruler are wrong with illogical position of some ticks, non logical sequence of numbers and unreadable numbers. A physical unit of length on ruler is missing. None of generated images could serve well in physics textbook. Figure 2: AI generated images for second prompt. Figure 3: AI generated images for third prompt. Conclusion In this paper, all AI generated pictures suffer from qualitative shortcomings in the area of physics. At first glance, presented pictures look nice, but they are more useful as unrealistic art or illustration for science fiction book rather than physics education material mostly because of AI hallucinations. Physics textbooks and publications must be based on images that truly describe reality of involved natural phenomena. Using AI generated nonsensical and hallucinogenic images in textbooks can lead to bad understanding of physics involved. It can be said that current state of AI for generation of images for physics education is very tricky and misleading, therefore teachers should be careful. One interesting application could be possibly the use of AI generated images as model puzzles for students to seek and find errors to test their already learned skills. References 1. G. Kortemeyer, "Could an artificial-intelligence agent pass an introductory physics course?," Physical Review Physics Education Research 19, 010132 (2023). 2. W. Yeadon and T. Hardy, "The impact of AI in physics education: a comprehensive review from GCSE to university levels," Physics Education 59, 025010 (2024). 3. A. Sperling and J. Lincoln, "Artificial intelligence and high school physics," Phys. Teach. 62, 314–315 (2024). 4. M. G. de Souza, M. Won, D. Treagust and A. Serrano, "Visualising relativity: assessing high school students’ understanding of complex physics concepts through AI-generated images," Physics Education 59, 025018 (2024).