It’s quite amusing what answers you might receive as they learn to classify their visual input according to our cultural standards. I sometimes

Color Naming: Human vs. GPT-4

submited by

Style Pass

2024-04-20 16:30:07

It’s quite amusing what answers you might receive as they learn to classify their visual input according to our cultural standards. I sometimes like to think of our current AI models as young children—but with rhetorical superpowers. They are eager to answer all our questions, regardless of knowing if they are right or wrong.

So, a new idea was born: Let’s ask ChatGPT to tell us what color it sees and compare it to human color naming in a small-scale experiment.

For this purpose, I used the GPT-4 Turbo with Vision API to run an experiment, showing the model more than 1,000 images, each in a different color.

I’ve then compared the responses from GPT-4 with the results of the Color Survey conducted in 2010 by Randall Munroe, the author of the popular web comic xkcd.

This might seem like a mundane result, but if we are to build AI systems that align with human understanding, it’s reassuring that we won’t have disagreements over color names.