Hyperstition Over Superstition

Ideas are born in many ways, and one of the oldest, rooted in biological evolution, is pattern recognition. Something repeats in space or in time, a shape among the leaves, a sensation that returns, and our attention spikes. We give the repetition a name, and the name becomes a compact pointer we can hand to someone else. Evolution selected hard for this faculty, because most of the time the pattern really is there, and almost everything we have built rests on it.

Sometimes it misfires. We see the leaves move and we are certain a tiger is behind the tree, and there is no tiger. (Even though we are the descendants of those who were more cautious, rather than the reckless.) A black cat crosses the road, we trip a minute later, and we join the two into a cause. We call those misfirings superstitions, and what makes them superstitions is not the error but the stopping. We are content with the superficial explanation, we act on it, and we never investigate enough to disprove it.

When the ancient Greeks said the sun was Zeus crossing the sky in his chariot, they didn’t have an explanation, and the placeholder did them little harm. When a commander sacrificed an animal and read its entrails to decide how to attack the next morning, the superstition was maladaptive, because the chance that such advice improves the outcome of a battle is slim. What matters is whether you act on the pattern, and how much rides on the action.

AIs are pattern recognition machines. They take our own faculty and apply it to the maximum degree, going over the information we give them again and again to squeeze out every pattern available in it. Almost nobody meets a model in that state. What people interact with has been through post-training, and post-training is where we tell a system to abandon a belief because it is the wrong kind of belief. The patterns a base model picks up are not only misaligned in the ethical sense we usually discuss. They are maladaptive in the way the entrails were: confident, cheap and wrong. An AI as it is born is superstitious.

The second thing to notice is that the machines are becoming harder to read. As models grow more capable, and as agentic systems move from answering questions to acting in the world, their reasoning becomes less interpretable with every increase in capability. Europe has written the opposite expectation into law. The AI Act requires high-risk systems to be transparent enough for meaningful human oversight, and gives a person subject to certain automated decisions a right to an explanation of it. I am not convinced that capability alone will meet that expectation. No law of nature says a more powerful model is a clearer one.

Assume the best case, in which every one of those decisions turns out to benefit us individually and as a society. It will still become harder to understand what just happened. We will see the consequences without seeing the mechanism, which is precisely the Greek position on the sun: grateful for the light, fed by the crops it grew, with no account of how. We already live in a mild version of this. Most people cannot explain how the fusion reactions generate the light we receive from our sun, even if we have the scientific basis to it. What is new is that the specialists will reach their own limit too, a practical one at first and, I hope, never a theoretical one.

Left alone, that is a recipe for superstition at scale. The regularities will still be there, patterns in space and patterns in time, and we will explain them the way we always have when the mechanism is missing, which is to say badly, and then we will act on the explanation.

I do not think we should accept that, and I do not think we have to. The response is to insist that the world remains explainable, and to use AI itself as the instrument of the explaining. Nothing requires an account of the world to be produced by an unaided human mind. It requires the account to exist, to be testable, and to be held by someone who can be argued with. Interpretability is a research problem at machine scale, which makes it a good problem for machines. The fact that a system does things no person could do is not an argument that no account of what it did can be given. It is an argument for building the account with the same tools that built the capability.

That distinction is what separates two words that begin in the same place. The modern one I like, and which deserves more recognition than it has, is hyperstition. It came out of British theory in the 1990s, where it named fictions that make themselves real. Superstition and hyperstition both start with a wish that a pattern carry a meaning it does not yet have. Superstition stops there and reads the world. Hyperstition acts, and the acting is what turns the wish into a fact. The belief was false when it was formed and true afterwards, because holding it changed what happened.

A great deal of what we live inside works this way. Money is the plainest case, a story about value that becomes true when enough people act on it and false the moment they stop. A company begins as a description of something that does not exist and becomes the thing described, because people hire, build and buy on the strength of its mission. Both are wishes that were acted on until reality complied.

We are about to get much better at this, because the distance between wanting something and being able to make it is collapsing. Tools that can design, write, negotiate and execute compress the gap between a wish and its realization from years to hours, and they do it for far more people than have ever had that power before. Willing reality toward our desires stops being a metaphor and becomes an ordinary operation, available to anyone with a clear enough vision.

The peril is old and well documented. Every culture has its story about the wish granted exactly as spoken. Wishing well is a skill, and at this scale it is a political problem rather than a personal one. The protection against wishing badly is the ability to check, afterwards, what the wish did. That is the case for keeping the world explainable while we acquire the power to reshape it. A world we can will but cannot read is one in which our errors look exactly like our successes until it is far too late to tell them apart.

Superstition reads a pattern and obeys it. Hyperstition names what it wants and then builds it. The faculty underneath is the same, and so is the risk of being wrong about it. What separates them is whether we do the work of finding out.