Pattern-Hungry: Why Humans Love Stories

In the past two weeks I've published two blogs about stories that arose around blood type: the one piece of biology we all happen to know about ourselves. This got me thinking about why humans love stories so much, and led to this post.

During the V-1 bombing campaign of 1944, Londoners became convinced the Germans were sparing certain neighborhoods. The bombs seemed to cluster, hitting some districts again and again while leaving others conspicuously untouched. The explanation people arrived at was that German sympathizers lived in the spared areas, feeding targeting information back across the Channel. Wealthy neighborhoods, poor neighborhoods, everyone had a theory about who was being protected and why.

An actuary named R. D. Clarke ran the numbers. He divided south London into a grid of quarter-square-kilometer cells and counted how many bombs landed in each. If the strikes really were clustering, non-random, guided by something, the distribution should have looked lumpy: some cells hit repeatedly, others suspiciously empty. What he found instead was a close match to a Poisson distribution, the statistical signature of pure chance events scattered randomly across a fixed area. No clustering. No sparing. The bombs were falling exactly where randomness alone would put them, and the human brains living underneath that randomness were doing something brains cannot apparently help but do: finding a story in it anyway.

That's the whole essay in miniature. Here's the slower version.

The brain you have is a prediction engine, not a camera

 It's tempting to think of the brain as something like a recording device, taking in the world and producing an accurate internal copy of it. That's not what's happening. Your retina, skin, and cochlea are together delivering something on the order of eleven million bits of sensory information per second. Conscious awareness processes somewhere around forty to fifty. The gap between those two numbers isn't a rounding error; it's nearly the entire signal. Almost everything reaching your senses gets thrown away before you're ever aware of it.

What survives isn't a copy of the world. It's a prediction, generated by an internal model that gets updated only when incoming data disagrees with it strongly enough to be worth correction. Consider your blind spot: there's a patch on each retina with no photoreceptors at all, roughly the width of your fist at arm's length, and you have never once seen a hole there, because the visual cortex fills the gap with a prediction drawn from the surrounding pattern rather than showing you nothing. Or consider phonemic restoration: erase a single sound from a recorded word, replace it with a burst of noise, and play it back. Listeners report hearing the missing sound clearly, correctly, with no sense that anything was ever removed. In both cases the brain isn't a passive receiver waiting for complete information. It's actively manufacturing the missing piece and presenting the manufactured version to you as though it were simply what's there. Researchers working in predictive processing, Karl Friston, Andy Clark, and Jakob Hohwy among the most influential, have spent the last couple of decades formalizing this: perception is mostly inference, corrected only at the margins by the world.

The reason for this architecture is not mysterious. A brain that waited to react until it had fully perceived a threat would be a brain that reacted too late. Prediction is faster than perception. So evolution built brains that guess first and check the guess against reality only as needed, because the guessing is what keeps the animal alive long enough to guess again.

Pattern-detection with no off switch

 Here's where it gets interesting. The machinery built to predict predators, weather, and the location of food doesn't come with a governor that shuts it off once survival needs are met. It's a general-purpose pattern detector, tuned by evolution for a specific set of adaptive problems, but with no way of distinguishing “a pattern that matters for staying alive” from “a pattern,” full stop. Once you have a system that scans constantly for structure and rewards you, neurochemically, for finding it, that system does not stay confined to its original job description. It runs on everything you point it at, including things with zero survival stakes.

You can watch this same circuitry misfire on something with nothing at stake at all. Sit at a roulette table and watch black come up five times running, and something in you will insist red is “due.” The wheel has no memory. Each spin is independent of the last, governed by exactly the kind of process that produces Poisson-style clumping over the long run, the same statistical signature Clarke found in his bomb data. But a brain built to assume that streaks mean something, a predator is near, a food source is reliable, does not easily accept that five blacks in a row are simply what randomness looks like some of the time. A pattern-hungry brain looks at a clump, at a card table or anywhere else, and assigns it a cause, because assigning causes is the job, and the job doesn't come with a built-in check for whether a cause is actually there.

Run that same overshoot through a slightly different circuit, one built for modeling other minds rather than physical events, and you get agency-detection: the reflex to assume a “who” behind an occurrence rather than a “what.” It's a useful reflex when the rustle in the grass really is a predator. It's a much less useful reflex when it's applied to a losing streak at a card table, a rustle that's just wind, or a universe that doesn't owe you a reason. But the circuit fires the same way regardless, because the circuit was never built to check whether firing was warranted. It was built to fire fast.

What happens when the machinery overshoots

 Most of the time this overshoot is just noise, a superstition here, a false pattern there. But four of the more notable things humans do look a lot like this same machinery, run deliberately and at scale, on inputs where the stakes were never survival to begin with.

Art is expectation, set up and then either satisfied or productively violated. A melody creates a prediction about the next note; the pleasure of hearing it resolve, or the deeper pleasure of hearing it almost resolve and then swerve, is the prediction engine registering its own model being tested. This is true even of art that refuses to resolve at all. A plotless novel, an atonal piece of music, a painting with no clear focal point, these aren't failures of pattern-making. They're a different kind of pattern, one where the recognizable shape is the absence of a shape, and the brain does the same sense-making work either way. It just makes the unresolved feeling the point.

Comedy runs on the same clock, just tuned differently. A setup builds a prediction about where a sentence, or a situation, is headed. A punchline breaks that prediction and replaces it with a second interpretation that also, on half a second's reflection, makes sense of everything that came before. The laugh is what a violated-but-resolved prediction feels like from the inside, which is why the same joke told too slowly falls flat: the first model gets too much time to settle in, and the swerve reads as merely wrong rather than delightfully wrong, and why a joke told too fast doesn't land either, since there was never enough of a prediction built to violate in the first place. Timing isn't an aesthetic flourish sitting on top of comedy. Timing is the mechanism, precisely dosed.

Conspiracy theories are the same causal hunger, unsupervised. This is, in miniature, exactly what happened over London: a real pattern of scattered, random events, and a population of brains reflexively assigning it a hidden intentional cause rather than accepting chance. Conspiracy theories run on that same logic at any scale, a cluster of events that feels too coordinated to be accidental, an agent posited to explain the coordination, a community that reinforces the pattern by trading confirming details back and forth until the story feels load-bearing. Agency-detection, the reflex to assume a “who” behind an occurrence, does a lot of the work here, and it costs the system almost nothing to overuse: assuming agency where there is none is cheap, while failing to notice agency where it actually exists can be fatal, so evolution built us to err generously toward “someone did this.” The reasoning isn't stupid. It's the ordinary causal-inference machinery, running exactly as designed, with no external check on whether the inferred cause is real. There's no separate gullible circuit sitting apart from the sensible one. There's one circuit, and it doesn't come pre-labeled with which of its outputs to trust.

Science is the same hunting instinct, except with a leash on it. A hypothesis is a prediction like any other, generated by the same overactive pattern-finder that saw spies in a bombing grid. What makes science different isn't that scientists have quieter versions of this instinct. It's that the discipline built an external check the brain doesn't supply on its own: test the prediction against reality, and be willing to let reality win. Clarke's own paper is an act of this discipline turned back on the very phenomenon that produced it, using statistics to catch the brain in the act of pattern-hunting and correct it. Science is, in that sense, the prediction engine's most successful attempt to police itself.

The part that doesn't resolve

 Here's the uncomfortable symmetry. The same system that produces a Beethoven symphony produces a conspiracy theory. The same system that let Clarke catch the brain in the act of pattern-hunting was doing the pattern-hunting in the first place. A cluster appears, therefore something must be causing the cluster, therefore let's find out what, or who. Sometimes that search turns up something real: a virus's mode of transmission, the actual cause of a bridge collapse, a genuine regularity in orbital motion. Sometimes it turns up a spy network that was never there. Most of the time, this machinery has no internal way of telling you in advance which search you're running.

I don't think there's a tidy place to land on this. The pattern-detection that makes you capable of art, of asking unanswerable questions, of doing science at all, is the exact same machinery that will hand you a false spy network with total conviction if you don't build something external to check it against. We didn't get the beautiful, meaning-saturated version of human cognition and the paranoid, error-prone version as two separate features. We got one system, running exactly as designed, and it was never going to be able to tell the two apart on its own.