The Ghost in the Average

Every image a model generates is an average. A weighted, statistical mean of millions of pictures it has seen, slightly bent in the direction of your prompt. When I ask for 'a cathedral at dusk,' the model is not painting — it is remembering, collectively, on behalf of everyone whose photographs were ever scraped, tagged, and pressed into its hidden layers.

This makes generative art a strange kind of séance. The model speaks in the voice of an aggregate — a crowd of dead photographers, forgotten illustrators, and tourists who once held up a phone in Florence. Your prompt is the question you whisper across the table. The image that surfaces is what the collective ghost agrees to say.

And yet — every now and then — the model fails. Six fingers. A doorway that opens into itself. A face whose left eye remembers a different century than its right. We call these 'hallucinations,' but they are also the only moments when the average breaks. Where the model produces something no human has ever made, and no other model has produced before. A glitch is the closest thing to originality a statistical machine can offer.

This is why I find myself collecting the failures. The mistakes feel more alive than the successes. The successes are competent renditions of every cathedral at dusk that has ever been photographed. The failures are the first cathedral that has ever existed — strange, wrong, irreducible.

We chase the glitch because that's where the human ghost still lives. Not in the mean. In the moment the mean cannot hold.