Limits of neurointelligence: Why can’t AI read time?

Neural networks can generate videos, narrate lectures, and map the universe, but they struggle to accurately read analog clocks. According to a study from the University of Edinburgh (ICLR 2025), even the most advanced models, including GPT-4.1 and vision-language models (VLMs), correctly interpret clock faces in only 38.7% of cases. When faced with calendars, the performance is even worse: just 26.3%, as reported by Livescience.

The Clock Test

It seems simple enough—glance at a clock and determine the time. But for AI, this is a spatial task requiring basic orientation and mental modeling, skills that current models lack. AI systems misjudge angles, confuse clock hands, and even with repeated instructions, fail to correctly interpret images. They are particularly thrown off by clocks with Roman numerals or unconventional designs. Researchers note that AI often “guesses” based on visual patterns without grasping the abstract concept of “time.”

In Space Without a Body

The root of these failures lies in AI’s fundamental “bodilessness.” It lacks physical experience, visuomotor coordination, and an understanding of object positioning in space. Humans develop spatial skills from early childhood through movement, touch, vision, and physical interaction with the world. AI, however, is trained on millions of 2D images, without any sense of “left,” “right,” or “behind.”

AI systems cannot mentally “fold” a box, determine Edinburgh’s location relative to London when facing west, or handle basic geometric transformations. Their “spatial thinking” is merely statistical pattern recognition on a flat surface, not an internal understanding of the world.

The Illusion of Universality

Paradoxically, these same models excel at passing SAT exams and writing code, creating a false sense of universality. If a neural network can explain quantum mechanics or generate a novel, why not trust it to navigate a city or make medical diagnoses? The answer is simple: it doesn’t truly understand. It predicts, not comprehends.

This uneven development isn’t a bug but a reflection of how AI is trained. It learns from texts, images, and dialogues, not from lived experience. The result: a brilliant conversationalist that gets lost when faced with a clock or the concept of west.