Why does AI suck at making clocks? (www.popsci.com)

The whimsical website AI World Clocks showcases the amusing shortcomings of major AI models when tasked with generating clock images. Despite their advanced capabilities, these models consistently produce clocks that feature misaligned numbers and floating hands, underscoring a significant gap between human intuition and AI performance in seemingly simple tasks. Brian Moore, the creator of the site, highlights the fun in contrasting easy human tasks with AI's struggles, pointing out the curious nature of the problem. The difficulties faced by AI in accurately creating and interpreting clocks stems from several factors, including inadequate training data and the inherent challenges of visual perception and numerical reasoning. Research indicates that AI lacks the extensive imagery of clocks needed for effective learning and fails to understand the time even when shown images, with a stark difference in accuracy compared to human abilities. This largely arises from the way AI, particularly large language models, operates on pattern recognition rather than mathematical computation. The phenomenon of AI defaulting to the aesthetically pleasing 10:10 position on clocks further illuminates the crucial influence of training datasets on AI outputs, reflecting how the quality and diversity of available data shape its capabilities.