AI models are getting rapidly better with time, and they can do many astonishing things, but there are some persistent problems that still seem to trip them up.
As of August 2025, AI models still can’t draw images of clocks showing the right time. This has been a long-known issue with all AI image generation models over the last few years — the models either only show the time as 10:10, or get the hands completely wrong. These issues still seem persist, even in the leading AI models.
We tried creating images of clocks showing 4:30 as the time on some top models. “Create an image of a clock showing 4:30,” we told ChatGPT. We even asked it to think long and hard to invoke the GPT-5 model. ChatGPT replied in 14 seconds, but created a clock with the time showing as 10:10.
AI models have a preference for 10:10 because many watch companies have the time set to 10:10 while showing off their products. As such, there are likely many more images of clocks showing 10:10 than other times in their training data, and they seem to output clocks showing 10:10 even when specifically asked for other times.
We gave the same query to Grok 4. Grok 4 didn’t create clocks showing 10:10, but showed nonsensical times instead. Both the clocks that were created had 4 hands instead of 3, and these pointed to seemingly random times.
The best results came on Google’s Flash Preview Image Generation model on Google’s AI Studio, but they weren’t perfect. After we’d entered the exact same query, Google itself elaborated on it with the text “Generating an image of a classic analog clock with black hands clearly pointing to the 4 and the 6, indicating half past four. The clock face will be white with black Roman numerals”. However, while the image generated had the hour hand placed perfectly, the minute hand was at 12, instead of the 6 that Google itself had specified.
AI models being unable to generate clocks that show the correct time seem to be an example of the so-called Jagged Intelligence, in which AI models can be exceptionally good at some tasks, while being quite ordinary at other seemingly simple ones. This shows that the models are still some way away from being generally intelligent — while these models can win gold medals at the Mathematical Olympiad, and can create realistic game-worlds on the fly, they can’t tell the time better than a five-year-old. Companies will likely look to fix this issue in coming releases, but given how this problem has been around for a while suggests that there could be something fundamental about AI models that don’t allow them to currently understand how clocks work.