 |
| AI Systems Accurately Read Clocks Only 38.7% and Calendars 26.3% of the Time – Image Credit: Alamy |
From composing poetry to generating lifelike images, AI seems unstoppable — until you ask it to read an analogue clock. A recent study has thrown a surprising curveball into the conversation about artificial intelligence: despite mastering complex language and coding tasks, today's most advanced AI models struggle with basic timekeeping skills like reading clocks and calculating calendar dates.
Presented at the 2025 International Conference on Learning Representations (ICLR) and published on arXiv on March 18, this research reveals a fascinating — and concerning — shortcoming in modern multimodal AI systems.
Clock Confusion: AI’s Surprising Blind Spot
Researchers at the University of Edinburgh tested several cutting-edge AI models — 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 goal? To see how well these models could perform tasks most humans learn in elementary school.
The result: underwhelming performance. The AI models were only able to identify the correct time from a clock 38.7% of the time and calendar dates just 26.3% of the time. In other words, a coin toss would fare better in some cases.
Why Is It So Hard for AI to Tell Time?
According to lead researcher Rohit Saxena, the issue isn’t that the models don’t recognize clocks — it’s that they can’t reason about them spatially.
“Clock reading requires something different — spatial reasoning,” Saxena explained. “The model has to detect overlapping hands, measure angles and navigate diverse designs like Roman numerals or stylized dials.”
It’s not just about identifying an object as a clock — it's about interpreting it, and that requires visual-spatial logic that current AI lacks.
Failing the Calendar Test
AI also stumbled when challenged with simple calendar calculations. For example, when asked, “What day will the 153rd day of the year be?”, models often got the answer wrong — not due to a lack of arithmetic skills, but because of a fundamental flaw in how large language models "think."
“AI doesn’t run math algorithms,” Saxena said. “It predicts the outputs based on patterns it sees in training data.”
So instead of logically calculating the result, the AI models guessed based on examples seen during training — and when those examples are rare (think leap years or obscure date formats), they fall short.
The Bigger Picture: Pattern Recognition ≠ Reasoning
This study serves as a reminder that AI’s intelligence is fundamentally different from human cognition. While it excels at recognizing patterns in vast datasets, it struggles with tasks that require abstract reasoning or combining perception with logic — especially if those tasks are uncommon in its training data.
What seems "simple" to a person — like telling time — can be extremely difficult for an AI model that has never been taught to spatially interpret overlapping hands or understand month-to-week conversions.
What This Means for the Future of AI
The findings underscore the importance of keeping a human in the loop, especially in time-sensitive or logic-critical applications like scheduling, automation, or assistive technology.
It also suggests that future AI development must go beyond scaling up datasets and instead focus on improving logical and spatial reasoning capabilities.
“AI is powerful, but when tasks mix perception with precise reasoning, we still need rigorous testing, fallback logic, and in many cases, a human in the loop,” Saxena concluded.
Final Thoughts
As we integrate AI into more real-world scenarios, it's essential to recognize where its limits lie. This study is a stark reminder: don’t set your watch by AI — at least not yet.
Sources:
- arXiv Preprint: “Evaluating Clock and Calendar Reasoning in Multimodal Large Language Models” (2025)
- 2025 International Conference on Learning Representations (ICLR)
What do you think? Should AI be trusted with tasks like scheduling and time-sensitive automation? Let us know in the comments or share your thoughts with #UReadDigest.
Post a Comment
0Comments