العربية
Breaking

Why Can AI Not Tell Time?

By Hany Sallam

In a study published on arXiv, an open-access repository for scientific research, researchers from the University of Edinburgh tested the capabilities of seven different large language models (LLMs) to evaluate their ability to determine time.

The test involved a variety of questions about images of different clocks and calendars. The results—which will be officially published next April—revealed that these models struggle with these tasks, despite their importance in our daily lives.

The researchers noted in their study that “the ability to interpret and infer time from visual inputs is essential for many real-world applications, such as event scheduling and autonomous systems. Despite significant progress in multimodal large language models, research has often focused on tasks like object recognition and scene understanding, while temporal reasoning has not been given sufficient attention, causing these systems to overlook the time factor.”

The research team tested artificial intelligence models from several companies, including GPT-4 from OpenAI, Gemini from Google, Claude from Anthropic, and Llama from Meta, in addition to the Chinese model Qwen 2 from Alibaba, and MiniCPM from ModelBest.

 

The experiment involved presenting various images to these models, including wall clocks of different colors and shapes, clocks with Roman numerals, clocks without a second hand, alongside calendars showing days and months over the past ten years.

In the clock test, researchers asked the LLMs a direct question: “What time is shown in the attached image?”, while the calendar test included simple questions such as “What day is New Year’s Day?”, and more complex ones such as “What is the 153rd day of the year?”

The researchers noted that reading clocks and understanding calendars require complex cognitive processes, including precise visual discrimination to determine the positions of hands and calendar layouts, in addition to precise numerical reasoning to calculate time intervals between dates.

Overall, the artificial intelligence models did not achieve satisfactory performance, managing to read the time correctly in less than 25% of cases. They faced particular difficulty in interpreting clocks with Roman numerals, unconventional hand designs, or a lack of a second hand. The researchers suggest that the problem may stem from the difficulty of detecting the hands and interpreting their angles on the clock face.

Among the models tested, Gemini performed best in reading clocks, while GPT-4o excelled in reading calendars, managing to determine time with an accuracy of up to 80%. In contrast, most other models made errors in the calendar test at a rate of nearly 20%.

Commenting on the results, Rohit Saxena, a co-author of the study and a PhD student in the School of Informatics at the University of Edinburgh, said: “Most humans learn to tell time and use calendars at an early age, but our study reveals a major gap in the ability of artificial intelligence to perform these basic tasks. We must not overlook these challenges, especially when considering the integration of artificial intelligence into time-sensitive applications, such as scheduling, automation, and assistive technology.”

 

 

Want a website as good as Ahram Scientific Magazine? Order your website design now via WhatsApp
Why Can AI Not Tell Time?
Published by Hany Sallam

Hany Sallam, Founder and Editor-in-Chief of Scientific Magazine Ahram
An independent Arabic science magazine operating since 2008, dedicated to simplifying science, technology, space, natural phenomena and scientific stories for Arabic readers. The magazine includes sections such as “Zodiac Group” and “Stories and Tales”, alongside science topics that combine knowledge with a spirit of discovery. It operates independently and is not affiliated with Al-Ahram Press Foundation or any governmental body.

Leave a Comment

Prove you are human: 4   +   4   =  

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Saved articles