Take a moment to observe the photo below of me holding my son, and notice the variance of skin tones. His mother is part Chinese, so he has an olive undertone to his fair skin, whereas I have a pink undertone.

While shopping for baby clothes with my wife earlier today, I prompted OpenAI’s new flagship model GPT-5 model to analyze two photos: one of my face, and one of my son’s, and to sample one pixel that would best represent my skin tone, and one pixel for my son’s. The output was the pair of swatches below.

GPT-5 used a perceptual LAB color model designed to approximate how the human eye interprets color rather than using raw RGB averages. Despite the sophistication of the algorithm, the result is wrong in a way that is almost comical.

If I had asked it to think harder, or used a different model such as Claude 4.1 or Grok 4, would the result have been more accurate? Perhaps. But would the Goldilocks model be able to generate a color that represents how my skin tone appears?

The problem lies in the very request: sample the pixel that looks most like my skin tone. This task cannot be carried out using computation alone. Algorithms can compute averages, medians, or brightest values. But “most looks like” is a phrase rooted in lived experience, in the phenomenology of appearance. The probability distributions LLMs use to predict the next token can't access phenomenal perception. The models can't “see red,” or have the experience of seeing any color.

When I see my own skin, I don’t experience it as a hex code. I experience a presence “in the world,” a shifting interplay of light, undertone, and memory. To collapse all that into one pixel is to confuse representation with appearance.

My skin tone does not appear to me as a mere hex code. And if it did, it would not be #916a58.

The illusion of representation is mistaking statistical operations for perception itself. The tone of something cannot be reduced to a mere pixel. The map is not the territory.