I can’t stop thinking about the website AI World Clocks. The principle is simple: all the main AI the models available in the market need to encode a clock and you can see the results. The problem: they are all great disasters.
The numbers always seem to be in the wrong place and sometimes outside of the clock itself. The hands may or may not be in the correct position and sometimes float in space outside the clock. Even clocks that are quite nice look…off, somehow.
“Telling time is a very human thing, very easy for us to do, and something we learn at a very young age,” Brian Moore, the artist behind the site, told me in an interview. “It’s kind of fun and entertaining to turn that around, to see something that a human could do very easily that a computer can’t do.”
I kept this site open throughout the process of writing this article and I can confirm: it’s very funny. But why AI so bad at that?
Well, one thing to keep in mind is that the site limits all models to around 2,000 tokens to generate its clocks and uses the same prompt for all models. With unlimited computing power and a very specific prompt, you could get a better clock from an AI system. But the question remains: why is it so difficult for AI systems? The reasons lie in how AI systems work.
The AI ​​can’t tell the time
The AI ​​isn’t just bad at creating clocks; it’s also bad for reading them. A 2025 study Technologist Alek Safar suggests that humans are 89.1% accurate at telling time on analog clocks, while the top-rated AI is only 39.4% accurate.
This study only hypothesizes about possible reasons, but the potential explanations are all interesting. The first is that there simply aren’t enough images of clocks in the datasets for AI models to learn to tell time accurately. Another reason is that clock images are difficult to describe precisely using language, which is what large language models need to process them.
Another Study 2025 conducted by the School of Computer Science at the University of Edinburgh also revealed that all major language models struggle to understand the time when shown an image of an analog clock.
“Our results suggest that successful temporal reasoning requires a combination of precise visual perception, numerical calculation, and structured logical inference that current MLLMs have not yet mastered,” the study states.
As I said, none of these studies claim to completely know Why AI is not good at these tasks. There are, however, some interesting factors to consider, including the data sets that AI systems use to understand the world.
We must understand that large language models, the technology called “AI” in contemporary parlance, do not really do mathematics. This is counterintuitive, because we are used to thinking of computers as mathematical machines, but modern AI technology is based more on pattern recognition. Clocks are an interesting example of this at work. The systems, instead of calculating the angles or positioning of hands to indicate the time, attempt to guess the time based on pattern recognition. Which, come to think of it, isn’t that different from the way I personally tell time by looking at a clock: AI systems are just bad at it. And there are interesting reasons for this.
The 10:10 problem
Go to your image search tool of choice and type “watch,” then follow the time you see on the watch faces. You will quickly notice that a majority of analog watches are set to ten past ten (10:10).
Why this particular moment? Because marketing. Watch and clock sellers have long known that setting a watch to 10:10 makes it more attractive to potential buyers. A 2017 study published in Frontier in psychology suggests that this might be because the angles of both hands resembled a human smile. Another consideration is that at 10:10, the hands do not cover the logo, brand name, or any other complications like the date. Basically, it makes for an attractive photo and has become a standard in watch and clock marketing.
As a result, many images of watches and clocks on the Internet are set to 10:10. This means that a large portion of the clocks in AI datasets are set to the same time. Ask any AI system to draw you a clock and most of the time it will set it to 10:10, sometimes even if you ask for a different time. That’s part of why Moore ended up building his website with hilarious AI clocks.
“I asked an image generator to give me an image of a clock at a specific time, and it absolutely couldn’t do it,” he told me. “I was getting a lot of 10:10s, even though I was giving them a lot of specific incentives.” Moore is not alone here—at least one Reddit user noticed it while trying to generate clocks set to a specific time.
It’s just a small rabbit hole about clocks and watches, sure, but it indicates something in the data that AI systems have access to that may affect their capabilities. Another theory that comes up in discussions about this topic: drawing clocks is a common test for dementia, which means there are some very inaccurate drawings of clocks on the Internet.
The people who make AI systems don’t completely understand how they work, so a lot of it is guesswork. And that’s what makes the AI ​​Clock website so fun: it provides insight into how these systems work.
PopSci Outdoor Gift Guide 2025