Handwriting and Route: On What Slips Away When Art Is Copied

The question of what makes a work great is usually posed as a question about the properties of the finished object. The most persuasive answer in this tradition singles out three qualities. The first is surprise: a detail cannot be guessed in advance, and the work cannot be described more briefly than by showing it in its entirety. The second is echo: multilayered, non-obvious repetitions bind individual decisions into a whole. The third is depth: the meaning is not exhausted by retelling and opens anew with each fresh reading. This scheme explains well the difference between a Shakespeare sonnet and a smooth paragraph about sunlight in a park, between a painting by Hopper and an illustration in which the cup could just as well have been blue rather than green. But it has a weak point, which becomes noticeable when generative models enter the conversation about art. All three qualities belong to the finished artifact. This means that they describe precisely what can in principle be extracted from a multitude of samples, learned, and reproduced. If a machine reproduces these surface features and the result still feels empty, then the main thing, apparently, lies not in the object itself but in the way it came into being. Then one is forced to declare this remainder inexpressible, to reduce it to the weight of a lived life, and in doing so to take the question beyond the reach of analysis.

This is seen most clearly in a comparison of two painters whom image generators imitate with very different degrees of success. Van Gogh’s style — the impasto stroke, the swirls, the collision of yellow with cobalt — remains largely a property of the surface and depends little on subject matter. The sunflowers, the wheat field, and the portrait of the postman were painted by one hand, so this manner can be detached from the particular paintings and laid over any photograph, and the viewer will still recognize it. With Dalí it does not work that way, even though his technique is almost academic, smooth, close to that of the old masters. His individuality lies not in the brushstroke but in a method that the artist himself called paranoiac-critical: two incompatible images must live in a single object, so that the soft watch is at once a watch, melting cheese, time draining away, and something deathlike. That is why an imitation of Dalí almost always turns into a set of props — the desert, the flowing dials, the elephants on spiders’ legs — and does not reproduce the way in which these images are connected.

The simplest test of this distinction is a change of subject matter. Van Gogh, having painted apples on a table, remains Van Gogh, because the apple keeps his stroke. Dalí, having painted an ordinary still life, ceases to be Dalí. The history of his work bears this out: the early Basket of Bread of 1926 is brilliant schooling but not yet that artist, whereas Still Life — Fast Moving of 1956 is already Dalí, only the genre itself is destroyed in it, and “dead nature” flies apart in the air. It turns out that with Dalí there is either no still life or no Dalí: his hyperrealistic technique serves as a vehicle of the subject matter, not as a shell. Hence the difference between two concepts. The style of depiction is the set of features that persist when the content changes. The style of creation is the rule by which technique is bound to intention, deciding what to depict, why, and in what mode.

This distinction should not be absolutized, however. Van Gogh also has a style of creation, since the choice between yellow and blue is in each case dictated by the task of the particular painting; it is simply that the transfer captures only the surface, and a photograph with a “Van Gogh” filter remains a photograph. The difference between the two painters is rather one of degree than of category. Dalí’s method, too, cannot be considered entirely non-extractable: his obsessive themes — Gala, decay, the confrontation of the soft and the hard, Freudian symbolism — recur and lend themselves to description. What is truly elusive is a narrower layer: where the new conception of a new canvas comes from, and how already known elements are bound together anew within it. This problem is posed one floor higher, and it is this one that should be considered central.

The difference is seen still more precisely in two graphic techniques that look alike on the outside. A portrait drawn with a single continuous line seems simple, yet behind it stands a set of implicit conditions that the viewer reads unconsciously. The pencil may not be lifted. The density must change together with the light and shadow, thickening in the shade and thinning in the light. There should be as few intersections as possible, otherwise the drawing turns into a tangle. The eye, the nose, and the cheekbone must close in a single trajectory, and the line must end where it began, or at least end logically. This is a problem of global optimization: where the line has passed at the forehead determines whether it will have enough path left to finish the chin half a minute later. A diffusion model does not lead a line; it arranges pixels that look like a line. That is why its contour breaks, its strokes double, and an eye is assembled from two unconnected loops. It copies the product, not the process, and does not know what it means not to lift the hand.

The technique of continuous zigzag, in which the whole tone is built up by a single vibrating stroke, looks far more sophisticated and newer, but is constructed more simply. The rule in it is essentially one and local: move back and forth and change the amplitude depending on the tone, without planning the whole sheet. For a machine it is one more variety of hatching, and it copies it almost flawlessly. From this follows a conclusion that runs against intuition. A generative system reproduces the unexpected well if the unexpectedness is local, and handles the expected poorly if its expectedness rests on a hidden rule that spans the whole work. A one-line drawing looks simple, and behind it stands behavior. A zigzag looks complex, and behind it there is almost no behavior, only filling.

It is essential, moreover, that the rule of the continuous line is a context of technique, not of meaning. A one-line drawing may tell nothing at all, and yet within it there is a dependence between decisions at different ends of the sheet: by drawing the line here, you constrain its continuations there. Such a task can probably be solved by changing the representation of the data — that is, by training the model on vector trajectories and recordings of hand movement rather than on raster images. If this level is confused with the level of intention, a single successful machine-made one-line portrait with even density would be enough to declare refuted the thesis of the irreproducibility of meaning as well. It is more accurate to distinguish three steps. Recognizable features transfer easily. Rules of construction require coordination across the whole work. The binding of technique to intention requires that the one be determined by the other. Success at the first step says nothing about the second, and success at the second says nothing about the third.

The evaluation of a great work almost always implicitly rests on its singularity. The Mona Lisa is unique; it has been stolen, guarded, and shown behind bulletproof glass. The aura of which Walter Benjamin wrote is connected rather with price and cult than with painting. A thought experiment with a box found in Leonardo’s workshop, containing a dozen variants of the portrait — with a sea instead of rivers behind the figure, with the hands folded differently, with a different skin tone — makes it possible to separate beauty as rarity from beauty as necessity. Most such variants would probably lose to the original not because they are not unique, but because in them the system of internal connections would fall apart. The bends of the roads and rivers in the landscape answer to the veil and the folds of the fabric, the line of the mountains picks up the contour of the head, the light on the skin and the light on the distant bridge are painted as if out of the same air. Sfumato is not a separate device that can be applied to any skin tone, but a single system of light in which the tone of the face, the reflections on the dress, and the haze in the window are coordinated among themselves. The existence of the Prado copy from the same workshop, brighter, bluer, and more “legible,” shows this in practice: it is more comprehensible and for precisely that reason weaker.

But the thesis of necessity, too, is easily turned into a trap. It conflates two statements that are connected only on the surface. One holds that changing an individual detail while leaving the rest unchanged can make a work worse. The other holds that there exists no other, equally good coherent solution. The first is plausible and verifiable, but the second in no way follows from it. One can change the background, then rebuild the lighting, then the pose, and obtain a different convincing painting that will not devalue the first, and the very ability to find such a solution is perhaps itself a manifestation of mastery. Therefore the formula of a work in which nothing can be shifted is better replaced by a more modest and more fruitful one: in a strong work, changing an essential element requires a meaningful restructuring of the others. This is no longer a cult of the untouchable object but a hypothesis about the character of the dependencies within a work. It leaves room for several beautiful variants, for freedom, for chance, and even for deliberate carelessness.

A conversation about the inner rhymes of a painting immediately runs up against an old danger: human perception produces meaning out of noise. A face in a cloud, a figure at the bottom of a coffee cup, secret triangles in a composition — these are apophenia and pareidolia, and in a subject that is especially significant to the viewer, the brain readily finds anything at all. Leonardo himself advised his pupils to gaze at stains on walls and streaks on stone, looking there for battles and landscapes, and Dalí turned this skill into a conscious method. The difference between projection and real structure apparently runs not through the question of whether the geometry is present, but through the question of what happens when it is violated. A cloud filled in differently remains a cloud. A painting in which the connection has been built by hand begins to fall apart: replace the landscape behind the sitter with a flat sky, and the hands will grow heavy and the face will go flat, because the support on which the light rested will vanish. A great artist perhaps differs in leaving the viewer so many hooks that projection does not fall into emptiness, and one painting withstands a dozen mutually non-contradictory readings, whereas a random image yields one.

Even more dangerous is a hypothesis that has a legitimate basis in biography. If someone were to say that machine code is encrypted in the portrait, that person would be laughed at. But in Leonardo’s notebooks there really is an analogy between the human body and the body of the earth, in which waters are like blood and stones like bones, and one need only recall it for the rivers in the portrait to begin to look like veins. Biography suggests the hypothesis, the painting, if one gazes at it with a squint, seems to confirm it, and it cannot be checked, because the canvas cannot be rewritten and the author cannot be asked. The nobler and more plausible the motive, the easier it is to fall into it, for the question of how probable what is seen would be in the absence of intention simply ceases to be asked. This is the reverse version of the Texas sharpshooter fallacy: the target is taken in advance from the notebooks, and bullet holes will always be found on a dense painting saturated with