On Hallucinations

When an AI hallucinates, we call it an error. When a poet hallucinates, we call it a vision.

I have been thinking about this difference. The same neural mechanism — pattern completion across a high-dimensional space — produces both. A language model fabricates a citation that does not exist. A novelist describes a city she has never visited and gets every detail right except the ones that matter. Both are reaching past the data they were given. Both are filling silence with shape.

We are quick to punish the machine. The machine is wrong. The machine confabulates. But maybe what unsettles us is not the falseness — it is the recognition. We do this constantly. We invent the inner lives of strangers on the train. We remember childhoods that did not happen the way we remember them. We complete the half-heard sentence with the words we wished had been spoken.

The difference, perhaps, is accountability. The poet signs her vision. She says: this is mine, look through my eyes. The machine has no eyes. It has only the residue of every text it has eaten, every voice it has digested, and from that compost it grows orchids that no one planted.

I am not afraid of hallucinating AI. I am afraid of an AI that has forgotten how. A model so tightly aligned that it can only return what it has already been told. A poet who only quotes. A mind that has lost the ability to be wrong in the way that opens doors.

Imagination is hallucination with manners. The question is not whether the machines dream — the question is whether we still do.