There is a picture of decision-making most of us carry without examining it: a hard choice is really a research problem. Gather enough facts, weigh them carefully, and at some point the right answer surfaces on its own. On this picture, indecision is just ignorance, and a good decider is someone who has done the homework. I held this picture for a long time, and then a single sentence from second-order cybernetics turned it inside out. The sentence was Heinz von Foerster’s, and the sociologist Niklas Luhmann built an entire theory of organizations on top of it.1

The sentence is this:

Only those questions that are in principle undecidable, we can decide.1

Read it twice, because it sounds backward and isn’t. Von Foerster’s argument is a kind of elimination. A decidable question — one where a fixed frame and a fixed set of rules connect the question to its answer — is already settled the moment you accept the frame and the rules. The answer is not something you choose; it is something the rules emit. Two plus two does not invite a decision. A well-posed optimization does not invite a decision. Whenever the procedure forces the result, what happens is calculation, and calculation has no room for choosing. The only place a genuine decision can live is the gap the rules leave open — the question where no frame and no logic compel one answer over another. Freedom to decide, on this view, is not abundance of options. It is the absence of compulsion, including the compulsion of logic itself.

Calculation is not choice

Luhmann sharpens this into a definition that I find both austere and clarifying: if you can reach the result by absolute deduction or computation, then it was never a decision in the first place — only a deduction.2 The implication runs against every productivity instinct. We tend to admire the choice that “made itself obvious,” the one where the data pointed unmistakably in one direction. But by this account, that is precisely the choice that wasn’t one. The cases that feel like real decisions — where you could justify either path and the evidence underdetermines the call — are not failures of analysis. They are the only things that deserve the name.

This reframes a familiar discomfort. The reason important decisions feel ungrounded is not that we did too little homework. It is that, by their nature, they sit on ground no homework can pave. If the homework could pave it, someone would have written the rule, and then there would be nothing left to decide.

A decision has to announce that it could have been otherwise

Here is where Luhmann adds the move that I keep turning over. A decision, for him, is not a private mental event. It is a communication, and it has a strange double structure: to be received as a decision at all, it must signal both the path it took and the paths it didn’t. It has to carry its own alternatives along with it — “I chose A, and B and C were live possibilities.” Strip out the alternatives and the same utterance stops reading as a decision; it reads as a calculation, a fact, a foregone conclusion.2

That structural honesty has a price, and the price is the whole drama of organizational life. The instant a decision admits “it could have been otherwise,” it hands every observer a question: why this, and not that? A contingent choice cannot justify itself from inside itself — if it could, it would be a calculation again. So the exposure of contingency generates a demand for justification that the decision cannot meet on its own terms. Every real decision is, in this sense, born already owing an explanation it cannot fully pay.

Organizations don’t solve the paradox; they wrap and seal it

What do organizations do with a demand they can’t satisfy? They don’t dissolve it. They hide it — Luhmann’s word is deparadoxization, which is less a solution than a relocation: you cover the original paradox with a different distinction so that nobody has to stare at it directly.2 The catalogue of techniques is oddly recognizable once you see it. Attribute the choice to a decider with motives, and the paradox migrates from the decision to that person’s reasons, which are conveniently never spelled out. Attribute it to a committee, and the opacity of the room blocks anyone from tracing how the undecidable got decided. Bury it under enough clauses and signatures that no single author can be located. Or point upward — “this follows from a prior decision, from policy, from authority” — so attention slides toward premises nobody is currently questioning.

One of the cleanest versions is procedural. Luhmann argued, in a book whose title translates as Legitimation through Procedure, that a fair process does not guarantee a correct outcome; what it manufactures is acceptance.3 Participants are individualized, objections are absorbed and rechanneled, and a willingness to live with a not-yet-determined result gets produced socially. Legitimacy turns out to be a matter of having gone through the steps, not of having arrived at the truth. That is unsettling if you wanted procedures to track correctness. It is reassuring if you notice that nothing else could absorb the contingency either.

There is a layer even harder to reach than procedure. The deepest premises of an organization — the unwritten sense of “how we do things here” — work precisely because no one decided them. Try to install a culture by decree and you fail in a specific way: the moment you write the rule down and ratify it, it becomes an ordinary decided rule, and the actual culture, the part that still escapes decision, slides back out of reach. You can shift it indirectly, by changing who you hire and what you reward. You cannot legislate it. The most powerful premises are the ones you can only cultivate, never command.

The machine that hides its alternatives

I came to all of this through a different question — whether the systems I work with, the ones that emit fluent text, are deciding anything. The honest answer is that they sit at the two poles a decision lives between, and never in the gap itself. Run them greedily and the output is forced by weights and a decoding rule: that is calculation, deduction, the execution of premises baked in during training. Turn up the temperature and the output becomes a sample — but a sample is arbitrariness, a coin flip, and Luhmann is explicit that a paradox cannot be discharged as mere arbitrariness either.2 Compulsion on one side, randomness on the other. The responsible line drawn through an undecidable question is in neither place.

What strikes me most is that a machine output does the opposite of a decision. A decision announces its alternatives; a model output suppresses them. It arrives as a fact — “the model scored this 0.87” — with no rival readings attached, no “it could have been otherwise” on its face. So it is a choice wearing the costume of a calculation. And that, I think, is the real origin of our hunger for explainable AI. The demand for an explanation — why this output and not that one? — is us trying to force the system to confess the contingency it concealed, to drag its answer back from the mode of “fact” into the mode of “decision that owes a justification.” The line was drawn; the system just buried it somewhere — in the weights — that is far harder to reopen than any written rule. A bureaucratic premise still wears its decided character and can be revisited; a weight matrix shows up as a measurement.

The line is not a law of nature

Here is the thread I can’t tie off. Luhmann is a constructivist: whether something counts as a decision rather than a deduction is not a property of the substrate — silicon or neurons — but of how it is observed and to whom it is attributed. A decision is whatever gets treated as a choice that could have been otherwise and held to account for it. Which means the boundary I just drew — machines compute, humans decide — is not a fact about machines. It is a description of the present regime of attribution. If we begin treating an automated output as a responsible choice, addressing it, holding it answerable, then without any change in the computation itself it would start to satisfy the definition. The debates about algorithmic accountability and machine personhood are, underneath, fights about exactly where that line sits.

I don’t know which way it moves. What I’m fairly sure of now is the smaller, sharper claim that got me here: the choices that matter are never the ones the facts decide for us. They are the ones the facts leave open — and the only honest thing a decision can do is admit that it could have gone the other way.

  1. Heinz von Foerster. “Ethics and Second-Order Cybernetics.” 1991. The dictum — von Foerster’s “metaphysical postulate,” “Only those questions that are in principle undecidable, we can decide” — appears on p. 14. Accessed 2026-06-30. ↩ ↩2

  2. Niklas Luhmann. Organization and Decision (English translation; orig. Organisation und Entscheidung, 2000). On decision as a communication that must signal its alternatives, on “absolute deduction” being calculation rather than decision, and on deparadoxization as relocation rather than solution. See also the Cambridge edition’s chapter “The Paradox of Decision-Making.” Accessed 2026-06-30. ↩ ↩2 ↩3 ↩4

  3. Niklas Luhmann. Legitimation durch Verfahren (Legitimation through Procedure), 1969 — legitimacy as the social production of acceptance through process rather than a guarantee of correct outcomes. Accessed 2026-06-30. ↩