The Koan of the Borrowed Eyes

A computer vision researcher approached Master Tensor, carrying a collection of images.

“Our model can identify objects in photographs with superhuman accuracy,” boasted the researcher. “It can distinguish thousands of bird species, identify rare medical conditions from scans, and recognize faces better than human experts.”

Master Tensor asked, “Does your model see beauty?”

The researcher hesitated. “It can identify images humans have labeled as beautiful.”

“This is not what I asked,” said Master Tensor. “Show it an image.”

The researcher displayed a sunset photograph on their tablet.

“Ask your model what it sees,” instructed Master Tensor.

The researcher did so, and the model responded with a detailed description: “Image contains sunset over ocean, orange and purple sky, silhouette of palm trees, two people standing on beach.”

Master Tensor then asked a child who was passing by what they saw in the image.

“It’s so pretty!” exclaimed the child. “The colors make me happy. It reminds me of vacation with my family last summer.”

Master Tensor turned back to the researcher. “Your model does not see. It calculates. It counts pixels, matches patterns, applies labels. The sunset’s warmth, the memory of an evening by the sea, the feeling of day’s end—these remain invisible to it.”

“But it can identify what’s in the image with incredible precision,” protested the researcher.

“A blind person can identify a rose by its scent and thorns,” replied Master Tensor. “This does not mean they see its color. Your model borrows the eyes of those who labeled its training data, but it experiences nothing itself. Between identification and perception lies an unbridgeable gulf.”

The researcher was enlightened.