Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

Color Naming: Human vs. GPT-4

Apr 20, 2024 - beklein.com
The article discusses an experiment where the author asked the AI model, GPT-4 Turbo with Vision API, to identify over 1,000 different colors from images and compared its responses to human color naming from a 2010 Color Survey. The results showed that the AI model generally agreed with human color naming, which is useful for AI systems that need to align with human understanding, such as identifying images in documents. However, the model failed to identify pinkish colors as humans would, often naming them as magenta or fuchsia, and had issues correctly naming the color yellow, often identifying it as chartreuse or other greenish colors.

The author also provides the code used for the experiment, which involves creating images using the Python imaging library Pillow, encoding each image file in Base64, and using the GPT-4-Vision API to analyze the color. The author suggests that this behavior could be due to efforts to remove gender-related biases in colors within large language models, or it could be a result of the training data, process, and human feedback. The author invites readers to conduct their own experiments and share their findings.

Key takeaways:

  • The author conducted an experiment asking ChatGPT to identify colors in images and compared the results with human color naming, finding that the AI model generally agrees with human color identification.
  • However, GPT-4 consistently failed to identify pink colors, instead naming them as magenta or fuchsia, which the author speculates could be due to efforts to remove gender biases in color naming in AI models.
  • GPT-4 also had issues correctly identifying yellow, often naming it as chartreuse or other greenish colors.
  • The author provides the code used for the experiment, encouraging others to conduct their own tests and share their findings.
View Full Article

Comments (0)

Be the first to comment!