The Missing Box
David Epstein opens Inside the Box with a story that sounds like a myth about genius. It is really a story about deadlines. Russian chemist Dmitri Mendeleev supposedly dreamed the periodic table into being overnight, a flash of pure insight. The truer version is duller and more useful. He had a chemistry textbook to finish, a shelf of known elements that refused to sit in any order he tried, and a publisher waiting. The famous “dream” was the output of months spent straining against a constraint he could not escape: describe every element, in some order, before the deadline.
That’s the thesis of Epstein’s book in miniature: humans do their best work when something pushes back. Bach’s fugues are extraordinary partly because a fugue is a rulebook. A melody must return, invert, overlap with itself, and still make sense. Michelangelo carved David from a block of marble that a previous sculptor had already ruined, leaving him almost no room for error. General Magic shows the same principle in reverse. This team of Apple veterans set out to build what amounted to a smartphone in 1990, with more talent, funding, and technical range than almost any startup in history. It folded after selling a few thousand units. Their founder later said he wanted to build engineers a kind of paradise, bounded only by what they could imagine. Without a defined customer, or a firm sense of what to leave out, the team could not agree on what to build. One engineer spent months expanding a simple calendar feature, first to cover two centuries, then all of human history, then all of astronomical time back to the Big Bang, because each new reviewer wanted it to cover one more imaginable case. A team with real limits would have shipped it in an afternoon. Unlimited resources dissolved General Magic’s focus.
The distinction that matters here is not whether a constraint exists, but whether it costs something. Emotion is largely what it feels like to hold a goal a person could fail at in a way that costs something real. Mendeleev’s deadline threatened his contract and his publisher’s patience. General Magic’s engineers faced no equivalent threat, only an unbounded budget and an unbounded timeline, and that absence of cost is what left them unable to stop.
Stravinsky made almost this argument in his lectures on music, decades before Epstein. The more an artist narrows the field of action, he said, the freer that artist becomes, because limitless possibility gives the will nothing to push against. The Oulipo writers took the idea to its extreme a generation later. Georges Perec wrote an entire novel, A Void, without once using the letter “e.” The missing letter generated the sentences, forcing substitutions and turns of phrase no unconstrained draft would have found.
The kind of limits humans can’t opt out of
What’s easy to miss in Epstein’s argument is that the constraints doing the real work are usually biological. Humans forget, so they invent writing, then libraries, then search engines. Humans die, so they compress a lifetime of care into a single letter, a single painting, a single conversation they don’t get to have twice. Humans get tired and distracted, so a strict word count or a hard deadline becomes a strange kindness. It ends the search for more options and forces a decision.
Herbert Simon, working the same territory from cognitive science, gave this a name: bounded rationality. Real decision-makers don’t scan every option and pick the optimal one. They satisfice within a limited search, stopping once something good enough turns up, because time and memory won’t allow anything else. Simon lived this on purpose, keeping his wardrobe down to a few interchangeable outfits and eating the same breakfast every day, on the theory that the best is the enemy of the good and that hunting for the optimal choice usually costs more than it’s worth. He pointed to what he called Fredkin’s paradox as the proof. People spend the most agonised effort on the least important decisions, precisely because the options are so close in value that no amount of comparison will settle it. This limitation makes deciding possible at all. Mihaly Csikszentmihalyi’s research on flow points the same way from psychology. People report their most absorbed, most creative hours when a task has a tight rule set and immediate feedback: a fixed chess position, a climbing route, a poem’s meter. The rules give attention somewhere to land.
In a complex, uncertain environment, a constraint prevents inquiry from continuing indefinitely, and action deferred indefinitely along with it.
Hemingway rewrote the ending of A Farewell to Arms thirty-nine times. Patrick Rothfuss reportedly rewrote The Name of the Wind so many times he lost count. Both had exactly enough patience and time to grind against a problem until something came out the other side.
The novelist Isabel Allende offers an even stranger version of the same discipline. She began writing almost by accident, drafting a farewell letter to her dying grandfather that turned into her first novel. Since then she has started a new book on the same date every year, clearing her office, going silent, and refusing paid work no matter how lucrative, for as long as the ritual demands. The only year she broke the pattern was after her daughter’s death, and even then the pull of the date eventually brought her back to the desk. The ritual functions the way a deadline or a fixed form does for anyone else. It removes the question of whether to begin, leaving only the question of what to write.
Where AI sits outside the box
This is the part where AI looks like the exception that proves the rule, and the idea shows up from a few different angles.
Call the first one the Freedom Fallacy: the common assumption that creativity requires total freedom, when research keeps showing that constraints sharpen creative output. A useful qualification sits inside this idea. AI can match average human creativity, yet doesn’t reach the top of human creative performance, and it can’t identify a problem worth solving or take initiative on its own. This is a different kind of limitation than the ones Epstein describes. It is an absence of the impulse that goes looking for a boundary in the first place.
There’s also a more structural way to frame the difference, borrowing from cognitive scientist Douglas Hofstadter’s Pulitzer winning Hofstadter’s Gödel, Escher, Bach. Human creativity thrives on paradox and self-reference: the strange loop of a person becoming who they are partly by making things, and being changed by what they make. AI stays bound by the shape of its own training and programming. It can imitate the form of a strange loop without living inside one. It can write a canon. Nothing about writing the canon changes the model afterward. There is no six months of scribbled attempts that leave a mark on it the way a poem leaves a mark on a person who spent half a year not getting it right.
Kant reached a version of this structural point first, from the other direction. In his account, the mind can only experience anything at all by imposing categories on raw sensation: space, time, causality, unity. These categories are the condition that makes perception possible in the first place. A mind without them would have no coherent experience to speak of. Wittgenstein made a similar argument about language in the Tractatus. The limits of my language, he wrote, mean the limits of my world. The wall is constitutive of thought, the shape thinking has to take to be thinking at all. Read next to Hofstadter, the three point at the same thing. For Kant, the categories. For Wittgenstein, language. For Hofstadter’s strange loops, self-reference. In each case the constraint is the architecture the mind is made of. An AI system has something with a similar function. A transformer’s attention mechanism imposes its own way of relating tokens to each other, structuring everything the model processes, in something like the way a fixed category would. It is probably a difference of degree rather than of kind. A transformer’s architecture is set at training time, while a person’s categories of experience can be stretched over a lifetime of new experience layered onto them.
The philosopher Bernard Suits argued directly against Wittgenstein on this point. Wittgenstein claimed that “game” has no common core, no single feature shared by chess, tag, and solitaire. Suits proposed that every game shares one thing: a voluntary decision to accept unnecessary obstacles, purely for the sake of the activity the obstacles make possible. Nobody has to dribble the ball, run the bases, or stay inside the lines of the court. Those rules make the game harder than it needs to be, on purpose, and that difficulty is the entire point. Suits extended the idea into a claim about a meaningful life generally. Value comes from freely choosing obstacles a person didn’t have to accept. Epstein’s Sabbath-like rules share this quality with Suits’s own attitude. An AI system doesn’t choose its constraints in this sense. Its limits are simply given. Suits’s whole point was that the choosing makes an obstacle meaningful.
A third angle turns on stakes rather than the constraints themselves. A prompt functions like a box. Nothing suffers if the AI ignores it or produces something forgettable.
This idea extends further. Meaning may not live exclusively inside individual human consciousness, and an AI system can already do a good deal of the functional work of interpreting a text or a situation. The case for human irreplaceability then rests on answerability: the capacity to be held responsible for an interpretation, to be changed by the experience of forming it, and to change someone else in turn. Mikhail Bakhtin’s idea of authorship turns on this point. To author a position, a reading, a judgment is to answer for having put it into the world. Emmanuel Levinas treated language itself as an ethical encounter, an address to another person that calls the speaker into responsibility before it does anything else. A student wrestling with a difficult text is authoring something in Bakhtin’s sense, taking a position they will have to answer for. Iain McGilchrist’s point about attention fits alongside this. What a person chooses to attend to is itself a moral act, since attention is what brings a thing into being for them. That responsibility does not transfer to a system generating an interpretation on request. AI output can be fluent, even insightful. It remains interpretation without answerability. Nothing is riskier for having said it, and nothing changes in the system for having gone through the process.
The people who will matter once AI can produce competent creative work on demand are the ones who write and question and make things for reasons an algorithm can’t hand them, because they chose to and stand to answer for it.
The asymmetry
These ideas point to an asymmetry. Humans carry involuntary constraints (mortality, fatigue, forgetting, scarcity of time) alongside voluntary ones layered on top, like a sonnet’s fourteen lines or a founder’s decision to ship with half the features cut. Epstein’s argument is that the voluntary kind works because it rehearses the involuntary kind. A deadline is a small, manageable rehearsal for the fact that nothing lasts forever, so a person had better finish the thing.
AI carries neither layer, or carries them in a form different enough that they barely deserve the same word. The precise version of this claim matters, because the loose version overstates things. AI is not free of constraints. Training data distribution, compute budgets, and reinforcement learning from human feedback all shape what a model produces, and RLHF in particular is a process built specifically to teach a system to prioritize some outputs over others, in a functional echo of what a deadline does to a person. A context window is a constraint too. But it costs nothing to hit it. The model doesn’t feel the running-out the way a person feels a deadline bearing down, and RLHF doesn’t cost the system anything the way a missed deadline costs a person sleep or reputation. The gap is not the presence or absence of constraints. It is that AI’s constraints don’t cost it anything, so they don’t press on it the way a real deadline presses on a person. Humans carry a dense, overlapping web of things that can actually be lost: livelihood, reputation, relationships. An AI’s goal is flat by comparison. Nothing downstream of its output can hurt it, so nothing downstream of its output is, for it, at stake.
A related frame, though a culturally specific one, comes from Sartre and Camus. Sartre argued that human beings are, in his phrase, condemned to be free: thrown into existence with no fixed nature, forced to choose and answer for the choice. Camus took the same starting point, a finite, mortal creature in an indifferent universe, and found in it the demand to make meaning anyway. Both describe a creature whose limits are inescapable, whose creative and moral life happens in the narrow space those limits leave open. A Buddhist or Stoic account of the human condition would frame this differently, so the existentialist language is not essential. The answerability argument above already suggests that since AI does not answer for its choices the way a mortal creature does, and it never inherits the pressure that, for Sartre and Camus as for Bakhtin, is where meaning-making starts.
An open question
This does not settle whether AI-generated work can be “good”. Plenty of it already is. It may explain a feeling many people report, that AI output is often smooth, competent, and slightly hollow in a way that’s hard to name. That feeling deserves a caveat. It may be a property of current statistical models rather than something inherent to any possible AI system, and a more advanced model might not produce work that feels hollow at all. If that is right, then smooth and competent AI output is being judged by the wrong standard when it is asked whether it is good. The real question is whether it can matter in the way a piece of work matters once someone has ground against a limit that actually threatened something. That is a harder question than whether the output is good, and probably the more important one.
None of this is necessarily permanent. Current training regimes are a contingent choice, not a law of nature. A system that had to revise its own output across multiple rounds, with a real cost attached to each round of revision, or that was penalised in a way that persisted and mattered to its future behavior, might develop something functionally closer to the pressure a deadline puts on a person. Nothing rules this out in principle. It simply is not how today’s systems are trained. So the interesting question is what a constraint would have to look like to cost the system something now, the way a deadline costs a person sleep.
There is also a collaborative version of this argument that deserves more room than a single line at the end. The most interesting cases right now are not AI replacing human creativity but AI extending it: a writer generating fifty variations on a line, a composer using a model to suggest harmonic moves they would not have thought of alone. In these cases the human supplies the constraint, the brief, the edit, the refusal to accept the first draft, and the AI supplies the fluency and the range of options. The resulting work can be more creative than either would produce alone, precisely because the human’s constraints are still doing the work of forcing decisions while the AI’s unboundedness opens up options the human would not have found on their own. This complicates the picture more than a strict human-versus-AI framing allows for, and it may be the more common and more useful version of the relationship going forward.
Two precedents suggest an answer. Tony Fadell, who led the iPod’s design and later co-founded Nest, treats constraint as something to manufacture when reality doesn’t supply enough of it. His team once built and argued over the thermostat’s packaging before the product existed, because deciding what earned a place on a box forced them to decide what actually mattered. His rule, roughly, holds that inventing constraints still forces the same sharpening a real deadline would. Brian Eno and Peter Schmidt built a deck of cards called Oblique Strategies on the same logic, to manufacture the kind of productive constraint a musician doesn’t naturally run into. The cards carry cryptic instructions like “honor thy error as a hidden intention” or “use an unacceptable colour,” drawn at random mid-session to break a stuck collaborator out of the default path. The cards work because they cost the artist something. A card has been drawn, it can’t be undrawn, and the piece has to reckon with it. This comes closer to what an AI system would need than simply shrinking its context window or capping its options, an imposed cost the system can’t route around, something it has to reckon with. The most interesting AI-assisted creative work right now comes from people who impose that kind of real constraint on the collaboration: tight briefs, hard edits, refusal to accept the first draft. They lend the machine some of their own scarcity, since it brings none of its own to the table by default.