Answers are grounded in this paper and cite their sources.
Introduction Large language models like ChatGPT are used in fields ranging from education to healthcare. This technology can simplify complex topics and help students study topics like anatomy. However, many concerns exist regarding its accuracy. This study aims to evaluate the image recognition and creation of DALL-E, ChatGPT’s native image generation engine, by examining its ability to label bones of the human foot and generate an accurate anatomical illustration. Methods This study utilized an untrained version of ChatGPT (version 4o). Two tasks were performed. First, ChatGPT was prompted to generate a labeled anatomical image of the human foot. The generated image was compared against anatomical references to assess its accuracy. In the second task, the model was asked to identify bones in three unlabeled foot images, two illustrations and one x-ray. Each image was inputted, with ChatGPT’s memory cleared between each prompt to prevent learning. The responses were compared to referenced anatomical labels, and accuracy was calculated for each. Results The study revealed inaccuracies in ChatGPT’s ability to generate accurate anatomical images and to identify parts of an uploaded image. While ChatGPT produced a visually detailed drawing, several labels were misplaced, critical bones were incorrect or not included, and entire structures were missing from the image. In the second labeling task, accuracy was low, with correct labeling rates of 27%, 57%, and 0% for each of the three image prompts. ChatGPT frequently misidentified and mislabeled bones. Discussion The findings highlight ChatGPT’s current limitations as an educational tool, particularly in its untrained state. The inaccuracies found underscore the challenges of using language-based AI models for specialized fields like anatomy. Factors contributing to these errors include ChatGPT’s reliance on text-based training data, which lacks the depth required for accurate visual representation. In addition to inaccurate data, the errors may also be due to ChatGPT not inherently having spatial or depth awareness – it understands patterns in data but not the geometry of real-world objects or figures. This study suggests that while AI holds promise as a supplementary educational resource, users must remain vigilant and always verify AI-generated content.
No takes yet. Share an insight, caveat, or question.
Shamith et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: