Edinburgh AI Clocks Study: Most Current AI Struggles to Read Clocks and Calendars

University of Edinburgh researchers have revealed a surprising limitation in current artificial intelligence systems: state-of-the-art AI models struggle significantly with basic timekeeping tasks such as reading analogue clocks and understanding calendars.

Published on March 13, 2025, the comprehensive study investigated the capabilities of multimodal large language models (MLLMs) to answer time-related questions by analyzing pictures of clocks and calendars. The results were startling, with AI systems correctly interpreting clock-hand positions less than 25% of the time.

The research team tested various clock designs, including timepieces with Roman numerals, different colored dials, and configurations with and without second hands. Mistakes increased significantly when clocks featured Roman numerals or stylized clock hands, suggesting fundamental issues with hand detection and angle interpretation.

AI models faced additional difficulties with calendar-based questions, including identifying holidays and calculating past and future dates. Even the best-performing AI model demonstrated a 20% error rate in date calculations, highlighting widespread issues in temporal reasoning.

Lead researcher Rohit Saxena from Edinburgh's School of Informatics emphasized the significance: "Most people can tell time and use calendars from early childhood. Our findings highlight a significant gap in AI's ability to carry out what are quite basic skills for people."

The combination of spatial awareness, context, and basic mathematics required to understand clocks and calendars clearly represents a weakness in current AI architectures. Unlike simple shape recognition, interpreting analog timepieces requires sophisticated understanding of angular relationships, numerical systems, and temporal concepts.

The research team found that AI systems performed no better when second hands were removed, suggesting the problems extend beyond visual complexity to fundamental issues with spatial reasoning and mathematical interpretation.

These findings have serious implications for AI integration into time-sensitive applications such as scheduling systems, automation platforms, and assistive technologies. As Aryo Gema, another Edinburgh researcher, noted: "AI research today often emphasizes complex reasoning tasks, but many systems still struggle with simpler, everyday tasks."

The peer-reviewed findings will be presented at the Reasoning and Planning for Large Language Models workshop at the Thirteenth International Conference on Learning Representations (ICLR) in Singapore on April 28, 2025.

This research underscores the need to address fundamental gaps in AI capabilities before broader real-world deployment, particularly in applications requiring reliable temporal understanding and basic mathematical reasoning.