Artificial intelligence dominates science but still struggles to read analog clocks

Artificial intelligence models now match or surpass human experts in scientific questions at the doctoral level and in math competitions. Yet these same machines struggle with a task that children learn early on: reading the time on an analog clock.

This contrast is highlighted in the AI Index 2026, an annual report from Stanford University published in April. The report calls this phenomenon an irregular frontier and cites a state-of-the-art model that correctly reads analog clocks only about 50% of the time, even while solving postgraduate-level physics problems.

Why is clock reading a challenge for AI?

Reading a clock involves several steps: locating the hands, distinguishing the hour hand from the minute hand, estimating the angle of each, and combining all this information to determine the time. A study by researchers from China and Spain tested multimodal models, which process both text and images, and found errors in reading, especially when the clock face or hands deviated from the standard.

Part of the explanation lies in the training data. Photos of clocks on the internet often repeat a few times, such as ten to ten. Without variety, the model may memorize patterns instead of learning the underlying rule, and the study itself questions whether machines have learned the task or simply memorized examples.

Researchers at the University of Edinburgh reached similar conclusions. The systems evaluated correctly interpreted the position of the hands in less than a quarter of the tests, and even the best model got 20% of calendar date questions wrong.

What do these errors reveal about the technology?

Uneven performance across tasks

Tests such as GPQA, which consists of postgraduate-level questions designed to resist simple searches, measure specific skills. A good result in this type of evaluation does not guarantee accuracy in visual perception or common sense tasks.

Impact on daily professional life

In practice, a tool can summarize an extensive contract while at the same time misreading an analog dial in a photo of equipment. Researchers at the Polytechnic University of Madrid point out that visual accuracy is critical in areas such as robotics, visual inspection, and assistive technologies.

For this reason, a common practice is to test the tool with its own use cases, define which steps require human review, and check numbers and images before using the result in decision-making.

The AI Index 2026 summarizes the situation: models have made progress in science and mathematics, but performance varies depending on the task. Knowing where each system’s strengths and weaknesses lie is what enables its safe use in the professional environment.