AI Struggles with Surprisingly Basic Tasks: Telling Time and Dates
San Francisco – As artificial intelligence continues to make headlines with its ability to generate complex text, create art, and even assist in scientific discovery, a new study reveals a surprising limitation: many current AI systems struggle with tasks that humans master in early childhood – reading analog clocks and interpreting calendar dates. The findings, stemming from research at the University of Edinburgh, highlight a significant gap in AI’s understanding of the physical world and raise questions about its readiness for real-world applications that demand precise time and date awareness.
The research, published as a peer-reviewed paper and slated for presentation at The Thirteenth International Conference on Learning Representations (ICLR) in Singapore on April 28, 2025, tested the capabilities of multimodal large language models (MLLMs) – AI systems designed to process both text and images. Researchers presented these models with images of clocks and calendars, posing questions that required understanding time and dates. The results were, frankly, underwhelming. The implications extend beyond a simple curiosity; reliable timekeeping is crucial for a growing number of AI-powered applications, from autonomous vehicles to scheduling assistants and tools designed to aid individuals with visual impairments.
The core issue isn’t simply recognizing shapes, according to Rohit Saxena, of the University of Edinburgh’s School of Informatics, who led the study. “Most people can tell the time and employ calendars from an early age,” Saxena stated. “Our findings highlight a significant gap in the ability of AI to carry out what are quite basic skills for people. These shortfalls must be addressed if AI systems are to be successfully integrated into time-sensitive, real-world applications, such as scheduling, automation and assistive technologies.” The challenge lies in the combination of spatial awareness, contextual understanding, and basic mathematical reasoning required to accurately interpret these everyday tools.
The Challenge of Analog Clocks
The study’s findings regarding clock reading were particularly stark. AI systems correctly identified clock-hand positions less than 25% of the time. The difficulty increased when clocks featured Roman numerals or stylized hands, suggesting that the models struggle with variations in visual representation. Interestingly, removing the second hand didn’t improve performance, indicating that the problem isn’t solely related to detecting a moving component, but rather a more fundamental issue with understanding angles and hand interpretation. This suggests a deep-seated issue with how AI perceives and processes spatial relationships.
This isn’t merely an academic exercise. Consider the potential impact on autonomous robots operating in a physical environment. A robot tasked with performing a time-sensitive action – delivering medication at a specific hour, for example – would be severely hampered by an inability to accurately read a clock. Similarly, a scheduling assistant relying on visual cues from a user’s calendar could make critical errors if it misinterprets the date or time of an appointment. The University of Edinburgh researchers emphasize that overcoming these limitations is essential for building truly reliable and useful AI systems.
Calendar Calculations Prove Difficult Too
The challenges weren’t limited to analog clocks. When presented with calendar-based questions – identifying holidays or calculating dates – even the best-performing AI model got the answer wrong approximately 20% of the time. This suggests that AI struggles with the conceptual understanding of time as a linear progression and the rules governing calendar systems. The ability to accurately calculate dates is fundamental to many real-world tasks, from financial planning to logistical operations.
Aryo Gema, also of the School of Informatics, points out the irony of the situation. “AI research today often emphasises complex reasoning tasks, but ironically, many systems still struggle when it comes to simpler, everyday tasks,” Gema said. “Our findings suggest it’s high time we addressed these fundamental gaps. Otherwise, integrating AI into real-world, time-sensitive applications might remain stuck at the eleventh hour.” The researchers argue that a greater focus on “grounding” AI in the physical world – giving it a more robust understanding of basic concepts like time and space – is crucial for unlocking its full potential.
Why is this happening? The limitations of current AI models
The study highlights a broader issue with current AI development. Many AI models are trained on massive datasets of text and images, but this training often lacks the kind of embodied experience that humans use to develop a common-sense understanding of the world. While an AI can learn to associate the word “clock” with an image of a clock, it doesn’t necessarily understand the underlying principles of timekeeping. This is particularly true for analog clocks, which require an understanding of angles, spatial relationships, and the cyclical nature of time. The University of Edinburgh’s research underscores the need for AI systems to move beyond pattern recognition and develop a more nuanced understanding of the world around them.
The research team utilized multimodal large language models (MLLMs) for their tests. These models, capable of processing both text and images, represent a significant advancement in AI. However, even these sophisticated systems demonstrate a surprising lack of proficiency in basic time-related tasks. This suggests that simply increasing the size and complexity of AI models isn’t enough to overcome these fundamental limitations. A different approach to AI training and development is needed, one that prioritizes grounding and common-sense reasoning.
Implications for the Future of AI
The findings have significant implications for the development of AI-powered applications. As AI becomes increasingly integrated into our daily lives, it’s crucial that these systems are reliable and trustworthy. An AI that can’t accurately tell the time or interpret a calendar is unlikely to be a dependable assistant in many real-world scenarios. The researchers suggest that addressing these limitations could unlock new possibilities for AI in areas such as scheduling, automation, and assistive technologies. For example, improved time-reading capabilities could enable AI-powered tools to help individuals with visual impairments navigate their daily routines more independently.
The study also raises broader questions about the direction of AI research. While much of the focus has been on developing AI systems that can perform complex tasks, it’s equally vital to ensure that these systems can handle the simple, everyday tasks that humans accept for granted. This requires a shift in emphasis towards building AI systems that are more grounded in the physical world and possess a more robust understanding of common-sense knowledge. The University of Edinburgh’s work serves as a valuable reminder that even the most advanced AI systems still have much to learn.
Key Takeaways
- Current AI models, even advanced ones, struggle with basic time-related tasks like reading analog clocks and interpreting calendar dates.
- The issue isn’t simply about recognizing shapes; it’s about understanding spatial relationships, context, and basic mathematical reasoning.
- These limitations have significant implications for the development of reliable AI-powered applications in areas like scheduling, automation, and assistive technologies.
- A shift in AI research is needed, prioritizing “grounding” AI in the physical world and developing common-sense reasoning abilities.
The research presented at ICLR 2025 is a crucial step in identifying and addressing these limitations. As AI continues to evolve, it’s essential that developers prioritize building systems that are not only intelligent but also reliable and trustworthy. The ability to accurately perceive and understand time is a fundamental aspect of intelligence, and overcoming this challenge will be critical for unlocking the full potential of AI.
The next step for the University of Edinburgh team involves exploring different training methods and architectures to improve AI’s time-reading and date-calculation abilities. Further research is also needed to investigate how these limitations affect other areas of AI performance. The team plans to present their findings at the Reasoning and Planning for Large Language Models workshop at ICLR in Singapore on April 28, 2025, offering a platform for discussion and collaboration within the AI community.
What are your thoughts on this surprising limitation of current AI? Share your comments below, and let’s discuss the implications for the future of artificial intelligence.
Related reading