The Failure That Cannot Count

Suppose that during a fuel shortage, a seller offers drivers a ritual that supposedly makes each litre carry a car farther. Customers can read the gauge and odometer, but their trials are few, and each tank is consumed under different traffic, weather, loads, and engine conditions. When one tank lasts longer, the preceding gesture receives credit. When the next does not, the timing, performance, or purity is blamed. The crucial step is not believing without evidence; it is adopting a rule under which evidence can only confirm.

Under noisy feedback, scarce trials, and shifting conditions, a useful heuristic becomes a ritual when failure is converted into a demand for more exact obedience. Causal opacity alone does not do this. You do not need to inspect combustion to compare fuel additives under stable conditions, or understand a model’s internals to A/B test prompts. A technique allows repeated comparison to weaken or discard it. A ritual has no losing result: success proves the formula, while failure indicts the practitioner, the setting, or an invisible exception.

Artificial intelligence recreates the favorable conditions. Its outputs are visible, but they vary with sampling, hidden instructions, model updates, and context. Most users conduct only a handful of uncontrolled trials. An excellent answer follows “Act as an expert,” and the phrase enters office doctrine. It may genuinely help; it may merely have accompanied a lucky output. The difference can be tested, but lore avoids the test by explaining every miss: the role was insufficiently specific, the punctuation disturbed the model, or the request lacked ceremonial politeness. The prompt expands because the claim cannot contract.

The same pattern appears when a remedy meets a fluctuating illness, a trading rule meets a changing market, or a creator tries to please an opaque ranking system. Each case supplies a visible result, weak counterfactuals, and cheap excuses. Ritual grows fastest where experiments are costly and rationalizations are free.

More measurement can therefore coexist with more superstition. A precise display can intensify the urge to optimize without improving the ability to identify causes. The candles arrive not when the dashboard goes dark, but when it becomes precise enough to command attention and variable enough to keep every comparison arguable.