When working with large language models like ChatGPT, it's easy to be impressed by their versatility, from answering questions to generating images. But how well do they handle detailed, practical instructions, especially in image creation? I recently ran a simple yet revealing test, and the results were surprising.

The Objective

My goal was straightforward. Ask ChatGPT to generate an image of a mechanical clock showing 7:30 (19:30). Seems simple, right? But this wasn't just any clock. I wanted ChatGPT to illustrate a traditional mechanical clock face, not just display digital text.

The Method

I prompted ChatGPT with a clear instruction:

"Can you make an image of a mechanical clock showing 7:30? I want to print it and post it on my partner's laptop because she always works until late."

ChatGPT generated an image, but not the one I asked for. The clock displayed 10:08, a time that’s oddly familiar in stock images and clock illustrations.

I repeated my request, stressing:

"Ok, that’s cool. But can you make the clock showing it’s 7:30?"

Again, ChatGPT returned another image, but still not showing 7:30. I pushed further:

"Can’t you make a mechanical clock showing 7:30 instead of 10:08?"

Yet the response was consistent… And consistently wrong.

The Result

No matter how I phrased the instruction, ChatGPT couldn’t generate a clock face set to 7:30. Instead, the model defaulted to the standard 10:08 positioning — a common convention in clock advertisements and stock images, likely because it’s considered aesthetically pleasing.

What This Reveals

This experiment highlights an important limitation in current AI image generation:

Data Bias: AI models trained on large datasets often reproduce patterns and defaults they "see" the most. In this case, the prevalence of clocks set to 10:08 in training data seems to bias the outputs.
Instruction Following: While ChatGPT is generally adept at understanding textual instructions, translating those nuances into precise visual outputs (especially in areas where data bias exists) remains a challenge.

Conclusion

This isn't just about a clock. It's a reminder that even the most advanced AI systems have blind spots. When training data is skewed, or when fine-grained control isn't well-embedded, the AI may fail to follow seemingly simple instructions.

For anyone exploring LLMs for creative or production tasks, this is a useful caution: test your prompts thoroughly, and don’t assume perfection, especially in image tasks with precise requirements.

ChatGPT image generation: A clock, but not quite on time.