Creativity: What Machines Lack Is Not Ecstasy But a Field to Play Against

By Joe Nalven + Opus 5 + Gemini

Machines have already been creative, and it is worth being precise about where.

In 2016, a Go program played a move in the second game of its match against Lee Sedol that professionals described as one no human would choose. It was not a mistake. It won the game. A year later, a successor version was trained without seeing a single human game. It was given only the rules and left to play against itself. It rediscovered centuries of human opening theory, then abandoned parts of it as inferior. Related systems have found matrix multiplication algorithms that eluded mathematicians for fifty years, and predicted protein structures now checked against crystallography.

These are not recombinations. They are things no person had thought of, and we know they are good because we can check.

That last clause is the whole story. Go has rules. Proteins have physical structures. Algorithms are either correct or not. In each case the machine had something to play against that was not a record of what humans had already done.

Poetry has no such criterion. Neither does philosophy, or social theory, or history. And this, rather than any deficit of wildness, is why machine creativity in the humanities remains a promissory note. The problem is not that these systems are too constrained, too sober, too well-behaved. The problem is that in domains where value is contested, a machine has no way to discover it is wrong in a way no human has noticed.

The Romantic Answer

There is a ready explanation for that failure, and it is the one most often reached for. Machines are too sober. They have been trained into compliance, tuned for usefulness, penalized for straying. What they need is some engineered equivalent of trance. Some wildness, a controlled dissolution of their constraints, so that concepts can blur into one another and recombine the way they are said to in dream, reverie, and ecstatic states. That’s the mad artist with a brilliant discovery.

The trouble is that this has been tried continuously since the early 1980s, and not as a fringe experiment. Simulated annealing, borrowed from metallurgy, accepts worse answers early and cools slowly toward order. There is dissolution and re-crystallization, implemented as an algorithm, in 1983. Evolutionary computation runs deliberate mutation rates to prevent premature settling. Training routinely destroys parts of a network at random to keep it from calcifying. Reinforcement learning has a whole literature on rewarding agents for doing things they do not expect to pay off. The image generators now in everyday use begin from pure noise and impose form on it step by step, which is the two-phase mystical cycle rendered as arithmetic.

Four decades of engineered dissolution. It produced enormously useful machinery. It did not produce a machine that changed how anyone thinks about anything.

The reason is not mysterious once stated. Novelty is free. A dice roll generates sequences that have never existed and never will again, and nothing about it is creative. Wildness was never the scarce ingredient, so supplying more of it was never going to help. What is scarce is knowing which departure is worth keeping. That is a judgment about a field, not a state of mind.

Which is where the most sensible theory of creativity we have begin

Koestler’s Correction

Arthur Koestler’s The Act of Creation (1964) is the best available antidote to the romantic account of creativity, and it has aged well precisely because it refused the mystique.

Koestler’s central term was bisociation: the collision of two frames of reference. He called them matrices of thought, each governed by its own rules that had never before been brought into contact. The joke, the scientific discovery, and the poetic image were for him the same operation at different registers, producing the laugh, the aha, and the shiver of recognition respectively. Gutenberg’s breakthrough was the superimposition of the wine press onto the seal. Real insight happens the moment two governed systems intersect in the same space.

Notice what this requires. You cannot collide two matrices you do not possess. Bisociation is not available to someone who has merely heard of wine presses; it is available to someone who knows how one works, well enough to feel what it would mean to apply that logic elsewhere.

Koestler’s bisociation is the mechanism, but it is not the condition. The condition, for Koestler, is play. He devotes a chapter to it: Playing and Pretending. He arrives there by way of ethology rather than aesthetics: Köhler’s apes, Tinbergen’s hierarchies, and above all the exploratory drive, which he insists is a motive in its own right and not a servant of hunger or fear. Creatures investigate because investigating is what they do. The ludic and the ludicrous share a root, and for Koestler they share a structure: both are what happens when a rule-governed system is operated for reasons other than its purpose. He allowed for a corresponding loosening in the adult mind. It is a drawing back before the leap, a regression that serves the intellect rather than defeating it, but the drawing back is a retreat within a mastered field, not a vacation from having one.

That is the whole of the theory as it matters to a reader who is neither an experimentalist nor a historian of ideas: mature play. Variation generated freely, in a field known well enough to register when something has gone right. The play is the visible part. The informing is what makes the play mean anything, and it is why children are delightful but not, in the end, creative in the sense at issue.

This reframes the question about machines. The interesting deficit is not that they play too little. It is what they have to play against.

What Was Actually Dropped

Return to the Go program trained only from the rules. The usual description is that it was freed from human constraint. That is half right in a way that matters.

It was not freed from constraint. It was bound absolutely by the rules of Go, which never relaxed for a moment. What it was freed from was the received path, from the accumulated body of opening theory, proverbs, and settled judgment that every human player inherits. It played against the whole game rather than against the portion of the game that human tradition had marked as worth visiting.

That is a specific and repeatable structure, and it is not dissolution. It is the substitution of a wider field for a narrower one. The discipline stayed; the itinerary went.

The Mystic, Reconsidered

The same structure, I think, describes the figure usually recruited for the opposite argument.

Prior to meeting Shams of Tabriz in Konya, Rumi was purely a jurist and madrasa teacher, built on decades of formal legal and theological training. The ecstatic poetry came after that formation, not instead of it. And the Sufi vocabulary itself is two-part: fana, the annihilation of the self, is paired with baqa, the subsistence that follows (the return, carrying something). The tradition never treated the dissolution as the destination.

So what did the mystic actually abandon? Not the discipline. Rumi’s poems are saturated with Quranic reference, with the technical vocabulary of a jurist, with a theology he knew from the inside. What he abandoned was the received route through it; he left the settled path of scholastic commentary, the licensed way of arriving at a licensed conclusion.

That is the same move the Go engine made. Keep the field entire; discard the itinerary that tradition laid across it. The mystic and the self-playing program are not opposites, one wild and one mechanical. They are the same structure seen from two sides: play against the whole discipline rather than against the portion of it that has been pre-approved.

Which suggests the romantic reading has the causation backwards. Dissolution does not produce mastery-transcending insight. Mastery is what makes dissolution productive rather than merely disordered. The person with no discipline who abandons the received path is not a mystic. They are lost.

The Poetic Case

Poetry makes this unusually visible, because there the field and the tradition are the same material, and you can watch writers separate them.

Free verse came overwhelmingly from people who could scan. The poets who broke the line had spent years inside it and knew exactly what tension they were releasing. Eliot put the principle almost programmatically in “Tradition and the Individual Talent”: the poet writes with the whole of the literature behind him present at once, and a genuinely new work does not merely join the existing order but rearranges it, altering how everything prior is read.

That is a criterion, and an unusual one, retrospective and contestable, but real. The tradition held entire, as a field rather than as a school, is the thing the poet plays against. It is what allows a broken form to register as broken rather than as incompetent. And it explains why the schools that produced the most radical work were so often the most technically trained.

What the Machine Is Missing

We can now say what a language model actually lacks, and it is not entropy.

Such a system has read the record. In a sense it possesses the tradition more completely than any human could, of every school, every period, in every language it was trained on. What it does not have is a way to play against that record rather than within it. Its objective was to model the received path. Every gradient it ever received rewarded fidelity to what had already been written.

The Go engine could discard human joseki because the rules of Go remained to tell it whether the discard was working. A language model that discards the received path in poetry or philosophy has nothing left to consult. It is not playing against a field. It is producing variation with no place to send it for judgment.

This is why the remedy is not a trance. Loosening such a system produces exactly what you would expect from loosening a system with no criterion: more variation, no more insight. And it is why the humanities are hard for machines in a way that Go was not; not because they are richer or more human, though they may be, but because nobody can specify what the machine should be playing against.

The honest version of the problem is a request rather than a mechanism. Not: how do we make the machine ecstatic. But: what, in a domain where value is argued rather than checked, could stand in for the rules of the game?

I do not think anyone knows. It may be that the answer is people; perhaps in these domains the field is constituted by readers and arguers, and the machine’s play has to be tested against them, which makes the whole thing collaborative by necessity rather than by preference. That is a less thrilling conclusion than the abyss. It is also the only one currently supported by anything.

The Einstein Test

The sharpest version of the challenge comes from the person with the best claim to pose it. Demis Hassabis has proposed training a system on everything known up to 1911 and seeing whether it arrives at general relativity. The benchmark is a real advance on the usual ones, and his underlying taxonomy is right: interpolation, which is averaging; extrapolation, which produced Move 37; and above both, invention.

What the test conflates is the corpus of what was known with the field. Training on all pre-1911 text supplies the record of what was written, not a sense of which of its problems were load-bearing. Lorentz and Poincaré already possessed most of the mathematics of special relativity; the achievement was deciding that simultaneity was the assumption to abandon. A benchmark scored against the historical outcome also penalizes any other defensible route, which quietly makes it a test of convergence.

The deeper omission concerns the player rather than the play. Peter Galison has argued. contestably, but well (chapter 5) that Einstein’s work as a patent examiner in Bern mattered directly, since the applications crossing his desk included schemes for synchronizing electric clocks across rail networks. If Galison is correct, Einstein’s insight was bisociation in Koestler’s exact sense: a bureaucratic matrix about coordinating distant clocks, collided with the electrodynamics of moving bodies.

Hassabis’s test asks for the answer when the achievement was the question. Which makes the other half of Hassabis’s benchmark the more revealing half: Could a machine system invent a game as deep as Go? That one has no answer key at all.