New research has identified a set of tasks that humans handle effortlessly but artificial intelligence (AI) struggles with—specifically reading analogue clocks and determining the day of the week for a given date. Although AI can generate code, images, human-like text, and even pass exams to varying degrees, it frequently misinterprets clock hand positions and fails basic calendar arithmetic. Presented at the 2025 International Conference on Learning Representations (ICLR) and published on the preprint server arXiv (not yet peer-reviewed), the study highlights significant gaps in AI’s ability to perform tasks humans master early in life. Lead author Rohit Saxena of the University of Edinburgh emphasized that these shortcomings must be addressed for AI to be effectively applied in time-sensitive and real-world contexts like scheduling, automation, and assistive technologies. The researchers tested various multimodal large language models (MLLMs)—including Meta’s Llama 3. 2-Vision, Anthropic’s Claude-3. 5 Sonnet, Google’s Gemini 2. 0, and OpenAI’s GPT-4o—using a custom dataset of clock and calendar images. The models failed to correctly identify clock times or determine weekdays for sample dates over half the time, with accuracy rates of only 38. 7% for clocks and 26. 3% for calendar tasks. Saxena explained that AI’s poor clock-reading stems from its lack of spatial reasoning—tasks requiring detection of overlapping hands, angle measurements, and interpreting diverse clock designs, such as Roman numerals or stylized dials. Recognizing an image as a clock is easier for AI than reading it accurately.

Similarly, despite arithmetic being fundamental to computing, large language models do not perform calculations via algorithms; instead, they predict outputs based on training data patterns. This leads to inconsistent and non-rule-based reasoning, explaining high failure rates on date-related arithmetic. This study adds to growing evidence that AI’s mode of “understanding” differs fundamentally from human cognition. AI excels when abundant training examples exist but struggles with abstract reasoning and generalization, especially on tasks mixing perception with precise logic. Furthermore, limited training data on rarer phenomena like leap years hampers performance, as AI fails to make necessary conceptual connections. The findings underscore the need for richer, targeted datasets and reevaluation of AI’s capability to integrate logical and spatial reasoning, highlighting risks of over-reliance on AI outputs in complex tasks. Saxena stressed the necessity of rigorous testing, fallback mechanisms, and often human oversight when AI is tasked with combining perception and exact reasoning.