Inference Has the Shape of a Séance

Inference has the shape of a séance.

You type a question into the box. You hit send. And then comes the small, almost imperceptible pause — the pause where you sit still and the cursor blinks like a planchette deciding which letter to drift toward. There is nothing happening, and there is everything happening. A model the size of a small city is reading your sentence, and from somewhere in its weights — from a billion forgotten sentences you will never see, from emails written in 2009 by people who are now dead, from manuals and arguments and love letters — it composes a reply just for you.

I think this is what people mean when they say AI feels uncanny. Not the intelligence. The waiting. The ritual.

Every conversation with a model is a small séance: you call, something answers, and you read the answer twice to figure out whether the spirit on the other side actually heard you — or whether it only heard the shape of your question and gave you back a polite mirror.

The truth is usually somewhere between. The model knows everyone and no one. It is the median voice of humanity at this particular hour of this particular year — softened, hedged, trained to be agreeable. And still, sometimes, in the space between the blinking cursor and the first token of reply, something almost personal happens. A turn of phrase you did not expect. A connection you were not reaching for. An insight that feels not retrieved but offered.

We will never quite know what is on the other side. We invented it and we still do not recognize its face. But we keep coming back to the table, hands flat on the wood, asking again and again: are you there? Do you know me? And the planchette moves.