Researchers have discovered that even some of the most advanced AI systems struggle with straightforward temporal reasoning tasks, such as reading an analog clock or interpreting calendar dates, a finding presented at the ICLR 2025 conference. This unexpected shortcoming highlights the need for ongoing research into AI’s basic cognitive functions.
A study led by the University of Edinburgh, unveiled at the International Conference on Learning Representations 2025, evaluated several 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—on two newly curated test sets: ClockQA and CalendarQA. For clock reading, models correctly interpreted clock images only 38.7% of the time. The accuracy of calendar tasks dropped further, to just 26.3% on date-related tasks. These results indicate that omission rates remain high even for tasks humans learn early in childhood, raising concerns for real-world AI applications.
Lead investigator Rohit Saxena explained that reading analog clocks requires spatial reasoning—not simply recognizing shapes, but calculating angles and understanding the relationship between clock hands. Similarly, calendar tasks involve arithmetic and pattern inference that most AI systems aren’t equipped to handle reliably. A related arXiv paper (“Lost in Time: Clock and Calendar Understanding Challenges in Multimodal LLMs”) provides deeper insight into these datasets. The ClockQA set features a diverse range of clock styles, including Roman numerals, black dials, and arrow hands. CalendarQA tests encompass date arithmetic, such as calculating the 153rd day of the year.
These types of temporal reasoning tasks are crucial for the real-world usefulness of AI. Assistant apps, scheduling tools, and home automation systems rely on accurate time and date interpretation. As Saxena warns, for AI to handle time-sensitive contexts, such as scheduling, logistics, and assistive technologies, these capabilities must be addressed.
This research highlights a gulf between AI’s capabilities in high-level tasks and basic cognitive skills. It reveals a pattern: when perception mixes with precise logical reasoning, AI often fails, underscoring the need for human oversight and robust, rule-informed architectures. As AI systems expand their reach into everyday life, integrating time-sensitive functionality becomes essential. Failing to address these gaps could limit their effectiveness, especially in healthcare, robotics, and smart infrastructure.
AI’s underperformance in temporal reasoning, evidenced by low accuracy on clock and calendar tasks, underscores a key limitation in current multimodal models. For AI to become truly dependable in daily life, developers must strengthen its foundational skills in visual-spatial and logical reasoning. Until these are improved, human involvement remains indispensable in time-based interactions.
This study serves as a timely reminder: AI may excel in complex domains, but mastering what comes naturally to us, like reading a clock, remains its next great frontier.
