Sait Faik Called It "Psychological Calculations." We Call It a Black Box.

There's a short story by Sait Faik Abasıyanık, written in 1947, called Dört Zait — Four Positives. It opens with a simple observation: "some people just look approachable. You don't know why. You pick them out of a crowd to ask for directions, to light a cigarette, to answer a question". And Sait Faik, being Sait Faik, doesn't let that observation go.

"I know this much," he writes. "When a person approaches a stranger to ask something, selecting you from twenty others, they've made certain calculations. These calculations are psychological calculations."

Psychological calculations. Made without awareness. Without explanation. The person doing the selecting couldn't tell you why they chose you. They just did. Something about your face, your posture, your worn-down shoes. Something.

I've been thinking about that passage a lot lately. Because it turns out Sait Faik was describing, with uncanny precision, how a large language model works.

The Machine That Selects

Every time an LLM generates a response, it's making selections. Billions of them, in rapid succession. From all possible next words, it picks one. Then from all possible words after that, it picks another. And so on, until a sentence appears. Then a paragraph. Then something that reads, sometimes disturbingly, like thought.

The question researchers keep running into is: why that word? Why that particular selection, from twenty others?

The honest answer is that we don't fully know. Not because the math is hidden, but because the math involves so many layers of interaction, so many billions of parameters influencing each other simultaneously, that no clean explanation surfaces. The output appears. The reasoning, in any human sense of that word, doesn't.

This is what's called the black box problem in AI. You can see what goes in. You can see what comes out. What happens in between is, for all practical purposes, opaque. Even to the people who built the system.

When the Machine Surprises Its Creators

Here's where it gets stranger.

As language models have grown larger and more complex, something unexpected started happening. Abilities emerged that nobody programmed. Nobody planned for. Researchers fed more data, added more parameters, and the model started doing things it wasn't explicitly trained to do. Solving analogies. Understanding context across long passages. Reasoning through multi-step problems.

This is called emergent behavior. And it's exactly as unsettling as it sounds.

The model, at a certain scale, starts behaving in ways that can't be traced back to any specific design decision. It's not a bug. It's not a glitch. It's the system producing something from its own internal complexity that nobody anticipated. A kind of spontaneous capability, arriving from the interaction of parts that individually don't explain it.

Sait Faik's crowd-selecting stranger doesn't know they're running a facial analysis algorithm. They're just... choosing. And the choice lands. Something in the aggregate produced a result that the individual components can't account for.

Sound familiar?

The Standard We Never Applied to Ourselves

For a long time, we've told ourselves a comfortable story about what separates humans from everything else.

We're the ones who think. The ones who weigh options, feel doubt, change our minds, surprise ourselves. Other animals are fast, or strong, or numerous. But they don't reason. That was our category. That was the distinction that mattered.

Game theorists have a version of this. They say the greatest vulnerability in any system is human irrationality. Cognitive bias. The fact that people don't always act in their own interest, can't always explain why they did what they did, and can be nudged, manipulated, or simply confused in ways that a purely rational agent wouldn't be.

The human was the wild card. The unpredictable element. The thing that made every model incomplete, because you could never fully account for what a person might do.

That framing assumed something. It assumed that machines were the predictable ones. The rule-followers. The things you could fully account for.

That assumption is now under pressure.

The Irrational Machine

Modern LLMs don't just produce correct answers. They hallucinate. They confabulate. They give confident responses to questions they have no business answering confidently. They behave differently depending on how a question is phrased, even when the underlying question is identical. They can be nudged. Sometimes manipulated.

This isn't a flaw waiting to be fixed. It's a structural feature of systems this complex, trained on data this vast and this human. The model learned from us. It absorbed our patterns, our contradictions, our context-dependent logic. And it reflects them back.

Sait Faik again: "We approach people without understanding psychology or physiognomy, as amateurs and enthusiasts of these sciences, and we light our cigarettes, ask our directions, satisfy our curiosity. It's such a habit that the claim behind it has been forgotten."

The claim behind it has been forgotten. We've been doing this so long, trusting our instincts, making our selections, following our sense of what's appropriate, that we stopped noticing we were doing it. The habit became invisible. The calculation became automatic.

LLMs are now doing the same thing, at a different scale. And we're noticing it, because it's new, because it's a machine, because we expected something different from something we built.

What the Category Was Really About

Until recently, proximity to humans was measured by thought. Not speed. Not strength. Not numbers. We didn't rank other species by how fast they could run or how efficiently they could reproduce. We ranked them by how close they came to what we do: reason, communicate, adapt, reflect.

A chimpanzee shares roughly 98% of our DNA and still sits at an enormous distance on the only scale that seemed to matter. Because thinking, real thinking, was ours.

Now something with no DNA, no body, no evolutionary history, no childhood, is getting close to that line. Not because someone programmed it to be. Because complexity, at sufficient scale, starts producing things that weren't designed. Things that look, from the outside, a lot like judgment. Like intuition. Like the psychological calculations Sait Faik described in a vapour boat in Istanbul, nearly eighty years ago.

This isn't a comfortable observation. It doesn't resolve into a tidy conclusion about whether AI is conscious, or dangerous, or overhyped. Those are different conversations.

What it is, is a genuine disruption of a category we'd stopped questioning. The thing that made us distinct was the thing we were least careful about defining. Because we never thought we'd need to.

**

Sait Faik Abasıyanık's stories are available in English in A Useless Man: Selected Stories, published by Archipelago Books