Artificial Intelligence Failed a School Task

Despite the ability to generate codes, texts, and images, artificial intelligence has difficulty telling time on an analog clock and reading a calendar.

In particular, scientists from the University of Edinburgh tested several multimodal language models that process text and visual information. They found that these systems regularly make mistakes in basic tasks. The study was presented at the International Conference on Representations of Learning.

As the lead author of the study, Rohit Saxena, said, most people learn to tell time and use calendars from an early age. The results showed a significant gap in the ability of artificial intelligence to perform basic skills.

For the experiment, the team fed a set of clocks and calendars into several leading models: GPT-4o (OpenAI), Gemini 2.0 (Google), Claude-3.5 Sonnet (Anthropic), and LLaMA 3.2-Vision (Meta). The results were disappointing: many models were wrong more than half the time – in determining the time by the clock and even in determining the day of the week for a given date.

Recall that in China, an AI robot attacked engineers.

To be continued…