Mario Carpo

Publication History:

Introduction, in Matias del Campo, Neural Architecture. Design and Artificial Intelligence, 12-15. Novato, CA: Oro Editions, 2022

The text posted here is a preprint draft and it is different from the published version. Please only cite from copy in print

The ancients did not have theories on what we now call “the fine arts.”  Where the moderns formulated lofty aesthetic principles on the modes and functions of mimesis in painting, sculpture, and architecture, the Greeks only had some fancy tales (aka anecdotes, myths, topoi, or parables) telling how some deft craftsmen on some noted occasions managed to make stuff that looked peculiarly natural—i.e., that looked as if made by nature.  So we know for example that the famed Zeuxis once painted grapes that were so lifelike that birds came down to peck at them, and Apelles painted a horse so perfect that its image fooled even other horses.  My cat, today, won’t be moved by any photographic picture nor even by moving pictures of other cats, let alone of dogs; but back then the arch-rival of Zeuxis, called Parrhasius, once won first prize in a painters’ contest by drawing a curtain that Zeuxis himself was tricked into attempting to pull aside (on that occasion Zeuxis admitted defeat and came in second).

The same Zeuxis also figures in another, quite different story. Having been invited to a town in what would now be southern Italy to paint a picture of a goddess, in search of inspiration he asked to see some examples of local babes.  The town elders sent him a selected group of handsome young men.  Zeuxis protested.  He was then allowed to see some girls; but finding none of them quite to his taste, he retained five of them as models.  His painting was a fusion, a blend, or an assemblage of features taken from all five, and it met with great success—hence the lasting popularity of the anecdote. From the point of view of art theory however, and even of the theory of human knowledge, this seemingly innocent tale conceals a number of major theoretical conundrums.  If Zeuxis already had an idea of feminine beauty in his mind, why did he need to imitate any real-life model?  And if, on the contrary, he did not have an innate idea of beauty, how could he choose among so many incomplete manifestations of the ideal?  The conflict between realism and idealism in the arts has found different solutions throughout the ages, but from a more practical point of view, the technicalities of Zeuxis’s mode of artistic operation—the parsing, selection, and the reassembly of parts coming from many models—have equally invited and prompted a never-ending stream of theories and speculations.  Evidently, the artist would not have limited himself in that instance to just cutting and pasting a number of pieces, as in a jigsaw puzzle; he would most likely have had to rework, modify, and adapt some of the parts, not only to chip off the edges but also more generally to make them blend with one another and merge in a single, harmonious composition.  Hence the question: how much of Zeuxis’s operation was what we could call today a collage, and how much of it would have been some looser and more creative form of imitation—the work of a talented artist only vaguely and distantly inspired by some of his models, or sources?  Could people look at his finished painting and tell: see, these are Emily’s eyes, Peggy’s nose, and Nancy’s lips?  Or did he blend all of his sources in one transfigured, truly supernatural composition, where one would say: see, there is a certain undefinable something in this portrait that reminds me of Emily, and of Peggy, and of Nancy, but it’s hard to tell what, precisely, comes from each?  In classical art theory, this is where art would have equaled nature, because this is the way nature works: this is the way a daughter may look like her biological mother.

This is also where Zeuxis could have vastly profited from today’s GAN technologies, of the kind that Matias del Campo and Sandra Manninger describe in this book.  In fact, hard to say if by chance or by design, GAN technologies today appear to do exactly what Zeuxis—as well as all painters working in the classical mode for the last twenty-five centuries—always strived to do: first, extracting a certain number of common features (or attributes or predicates) from a set of carefully chosen, compatible visual samples (in the case of Zeuxis, the dataset consisted of Zeuxis’s own pick of Sicilian beauties); then, using this list of common features to generate a theoretically unlimited number of copies that will all be similar, to some extent, to each of the models, but identical to none (Zeuxis appear to have produced just one copy in the story as told, but the principle is the same).  Each copy will remind us, somehow, of some or all of the models, but no one will ever be able to say in what precisely, or why, or where, the models and the copies look alike. 

The mystery of creative imitation has been at the core of the classical theory of mimesis in the visual arts since the beginning of time.  And this is what GAN technologies (i.e., one the latest avatars to date of data-driven machine learning) are now starting to tackle.  Just like print and photography, good old mechanical technologies, mastered the art and science of making identical copies, and just like identicality (the making of identical reproductions) was a trope and tenet of modernism in the arts, today’s digital tools are inaugurating a new trend of mass-produced similarities: the making of copies that are never the same, but all have a certain something in common—thus they resemble one another and their models. 

In short, today’s AI is about to automate imitation.  And in true AI fashion, AI is doing so without telling us how that happens: creative imitation remains a mystery even when carried out inside the black box of a recursive algorithm.  Which common features are being extracted from the dataset we feed to the system?  And how are these common traits embedded in the new images we can derive from the original models?  In other terms, what do the items in the original dataset, and the newly created ones, have in common?  Why do they look similar?  Nobody knows.  In that, we are not an iota more advanced than any of our classical predecessors.

When dealing with what we would now call highly structured datasets, classical artists soon found another, equally generic way to summarize the loose ambit of visual resemblances among works of art that have something—a certain undefinable something—in common: since Vasari, it was understood that works made in similar ways (showing the “hand” of the same master, for example, or the distinctive traits of the same school of painting) might be said to be in the same “manner.”  In the course of the nineteenth century, the term “style” was largely adopted with a similar meaning.  In the classical tradition, all art was to some extent imitative; all artists had to learn the art of copying; some ineffable similarity between the model and the copy was the sign of art well done; artists that produced creative copies in equally ineffably similar ways were said to be working in the same style.

That was then.  The theory and practice of creative imitation, and the related critical category of style, were equally obliterated by the modernist leviathan.  Seen from the vantage point of late-romantic sensibility, all copy is plagiarism; from the point of view of modernist morality, all imitation is sinful; from the point of view of twentieth-century iconoclasm, all discussion of style is wasteful.  Did we ever stop copying?  Of course not, because that’s to some extent inevitable in all we do.  Did we ever stop noticing that some artworks happen to be in the same style as some others? Of course not—we just found other, more opaque, hypocritical and labyrinthine ways to say so. 

Matias del Campo and Sandra Manninger’s work, as documented in this book, powerfully emphasizes to what extent the use of AI today obliges as to reassess some aspect of our natural intelligence that twentieth-century industrial modernism had made us forget.  It was time.  Imitation and style are back to where they should be—not as judgments of value, but as indispensable creative and interpretive tools.  The work of Matias del Campo and Sandra Manninger also demonstrates to what extent our post-industrial, computational future is bound to be closer in spirit to our pre-industrial past than to our late-industrial present.  Given the deliquescent state of our late-industrial present, that gives us more than a glimmer of hope.  

(New York, August 2021)