Formal Analysis, Generative AI, and the Eternal Return of Precedent


Publication History:

“Formal Analysis, Generative AI and the Eternal Return of Precedent.” Log 58 (October 2023), 133-139

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

In the early morning of a foggy day in December 2022, I took a more than usually rickety train to New Haven to sit as a critic in the final review of Peter Eisenman’s advanced studio at the Yale School of Architecture, cotaught that term with Daisy Ames and Frank Gehry. I know that studio—and the formal analysis class to which it is tied—for having often attended both as an auditor, critic, or discussant during my years at Yale. As generations of students know, both classes aim at learning architecture from the study of a corpus of historicallyrelevant architectural models. Eisenman chooses those models so that each collection may serve as a testing ground to run his design theory at full speed. His design theory, in turn, is meant to be exemplified by his own analysis of the models he shows—a mode of analysis famously known in North American design schools as formal analysis.

It would take a great philosopher and art historian to find out, precisely, what “formal analysis” is and whence it came to us. For the more limited scope of this anecdotal note, it may suffice to say that the prime culprit for the rise of formal analysis in design teaching was Colin Rowe, who seemed to think that modernism was the continuation of classicism with different means, and liked to read modernist architecture (Le Corbusier, for example) using terms and categories that would have been current in a French academy of fine arts around the mid-19th century. Peter Eisenman, then Rowe’s colleague at the School of Architecture of the University of Cambridge, had a different take on the matter, starting with his own PhD dissertation, The Formal Basis of Modern Architecture (submitted in 1963, and now available as a book).[1] In his dissertation Eisenman agrees that the purpose of formal analysis is to provide a universal reading of the language of architecture, valid across all styles and times (and, we may surmise, places). Eisenman’s architectural language, however, speaks primarily to the mind of a sentient human body moving in space, hence it is a language of invisible voids determined by solid enclosures. In architecture, that spatial language was and still is a modernist language par excellence.

Much to the frustration of many generations of students, Eisenman never formalized the lexicon nor the operations of formal analysis, which his students to this day have to intuit, infer, and extrapolate from Eisenman’s writing and lectures. And indeed, a vocabulary of formal analysis, itemizing and defining all the terms and expressions involved in the process, would be an excellent subject for a crowdsourced scholarly project. Figure-ground, part-to-whole, massing, screen, cuts, layers, warp, rotation, diagram, vectors, centrifugal, compression, etc.: In spite of their apparent ordinariness, none of these terms is self-explanatory. In fact, one could argue that simply learning the lexicon of these spatial operations is the primary know-how that students can garner from the teaching of formal analysis.

That, however, would only be the first part of the story. Formal analysis teaches us to read a language of architecture, and this may be good enough for, say, an art historian, but not for designers, who may still want to learn how to write in that language. And on that, Eisenman does offer more precise guidelines—almost a method, set out in 2008 in his Ten Canonical Buildings.[2] Here, Eisenman seems to embrace some core tenets of modern structural linguistics, which long argued that all use of languages depends on the dialectics between code and creativity, convention and invention (Saussure’s langue and parole). In the artsy version of structural linguistics popularized in the 1970s by Umberto Eco and Roland Barthes, the amount of information conveyed by a message grows in proportion to its quirkiness (or statistical unpredictability), but only within limits: Creativity and invention must not hinder the legibility of the code on which a message is based; when that code is lost, so is all meaning, because no one will know which language we are speaking. In other words, breaking rules is only meaningful insofar as one can tell which rules are being broken; all revolutions presuppose institutions. And in architectural terms, that means all creation, or invention, must first acknowledge the tradition to which it refers and into which it inscribes itself. Only when that tradition is stated may contestation and transgression ensue – and, in Eisenman’s terms, criticism. In short: choose, then know, the precedent against which you take a stance. This precedent can be a single model or, more often, a corpus of models: a canon.

One may argue that sometimes we want to take a stance against something, and sometimes not, but that would be a subject for another conversation. In purely linguistic terms, it remains nonetheless true that an entirely predictable message (a message where all is conventional, without any trace of invention) is bound to go unnoticed; whereas messages where all is invention, without any regard for shared conventions, are likely to be dismissed as—literally— idiosyncratic. Communication appears to be more effectively conveying information when a message shows a creative use, or misuse, of some recognizable convention or precedent it embeds. Well before the rise of structural linguistics, and of modern information science, many art and language theories had already come to very similar – after all, rather sensible – conclusions. I trust that comparable ideas of precedent must have also surfaced and blossomed outside of the European tradition – that is a part of the story I do not know. In the European tradition, however, at one point in time – known today as the Renaissance – reference to precedent famously became a mandate, and almost an obsession: in the words of one noted polemicist at the time, a pervasive disease.[3]

For reasons too long to explain, and which some found and still find inexplicable, around the end of the Middle Ages some European scholars started to argue that one period in the history of Europe, the period we now call classical antiquity, was the climax and zenith of all civilizations—past, present, and future—and that modern scholars, scientists, and artists, and even society at large, should revive the lore of their classical predecessors by imitating them. Evidently, that improbable project soon opened a Pandora’s box of practical snags and theoretical conundrums. Renaissance scholars had to find ways to imitate classical Latin while expressing ideas that did not exist in classical antiquity; likewise, Renaissance architects had to find ways to imitate classical buildings while designing buildings that did not exist in classical times (Christian churches, for example).

And this is where Renaissance artists were more or less obliged to come up with a new mode of imitation: imitation 2.0, as we would say today. Renaissance artists learned to imitate the form of an artistic model—but only its form, without its content. Their new mode of imitation thus became, first inadvertently, then more explicitly, the transference of form through different contents: the extraction and replication of some appearances regardless of subject matter.[4] They did not have a word to define the object of this new operation, until Vasari started to call it a “manner.” Manner is an ineffable yet transferrable quintessence, a certain something (nescio quid) infused into, and common to, different artworks: It is a feeling, a flair, a generic, pervasive, diffuse but undefinable aura. In the 18th century the term taste was often used with similar meanings, and as of the 19th century the term style (Cicero’s own term, ironically) took over, and replaced all others. In this sense, and with these meanings, imitation and style became the backbone of academic training and the warp and weft of the classical tradition in the visual arts, including the arts of design.

This is what modernism brutally and thoroughly eliminated from the arts and design of the 20th century.[5] Modernist artists, designers in particular, had many good reasons to feel averse to the ideas of imitations and styles that had dominated 19th-century European art theory. Yet the founding fathers of architectural modernism did not jettison the notion of style altogether. In 1932 in America modernism itself famously became a “style”— an “international” one. An all-out crusade against styles seems to have been an ideological project of late modernist zealots, climaxing in the 1960s and ’70s. This is when all imitation came to be seen as identical copy, hence plagiarism, or larceny, and art theory and literary criticism had to bring in a panoply of synonyms, circumlocutions and euphemisms to refer to the same, now unmentionable, notional field: inspiration, influence (often associated with “anxiety”), sensibility, inclination, affinity, etc.[6] The history of the demotion of ideas of imitation and style from 20th-century art theory is a fascinating one, and it does admit a few exceptions – particularly in postmodern times, for evident reasons. But from the 1960s to this day, the one singular, consistent exception to the antimimetic drive of late modernists—the one vox clamans in deserto—came from the formalist precinct. Before the pomo tsunami, nobody dared to plead for the dialectical role of precedent in design except the formalists, and formal analysis was their primary means to that end. And it remained so, a stubbornly resilient modernist tool for the critical assessment of, analysis of, and abstract reference to, precedent—even in more recent postmodernist times of chunky cut-and-paste and showy historicism.

I was reminded of the singularity of the formalist project—and of the quixotic and almost heroic, enduring isolation of its primary pedagogical tool, formal analysis, when I noticed, on that early December day in 2022, that some of its traditional keywords and general ways of referring to precedent started to have an eerie resonance for many in attendance. Those were the days, it will be remembered, when ChatGPT was making headlines, prompting a spectacular and unexpected resurrection of artificial intelligence—itself an old idea, generally dismissed until recently as a failed dream of early computer science. A few months earlier, the release of the first text-to-image applications had produced a similar commotion in the visual arts, even though design professionals had been working with a very similar machine-learning technology called Generative Adversarial Networks (GAN) since the mid-2010s.

At the time of this writing (summer 2023) GAN art, or art generated by AI, is a staple of digital art, and its image-processing strategies are well known. First, GAN artists must gather a corpus of images seen as related; then, a machine-learning algorithm processes this “dataset,” looking for traits that these images may have in common and formalizing the commonalities among them. This inductive process results in a mathematical matrix that computer scientists call a “latent space,” and roughly corresponds to what philosophers in the past would have called the definition, idea, formula, or quintessence of the original dataset. This definition, which today is neither verbal nor visual, but mathematical (vectorial), will in turn be used to recognize the content of external pictures or to generate new pictures derived from the same dataset. So, for example, if a dataset contains images of dogs, the system will recognize a real dog when shown one, or create realistic images of fake dogs if asked – apparently out of thin air, but in fact mapping each new dog to all (and only) the dogs already known to the system. With further tweaks, the technology can be used to imitate two datasets at the same time, thus creating a hybrid or fusion from two corpora of images, or to extract some generic but distinctive features from the first dataset and infuse them within the second. Around 2016, a team of computer scientists – evidently familiar with Western art history—started to call this operation “style transfer.” The term stuck, and is now embedded in some of the most popular AI-based image-making tools.[7]

AI-based image making is, at its core, a double exercise in similarities—visual similarities—that GAN algorithms must find, then replicate. The first part of the machine-learning process is analytic: to return to the example mentioned above, the system must first figure out what all those images of dogs have in common. What visual features make a dog similar to all other dogs—hence, a nominal dog? This inductive process will create the (latent) definition of an archetypal dog. Then comes image generation: how many variations can we introduce in the image of this ideal dog so that it will be visually different from its archetype yet still recognizable as a dog, insofar as it is like all dogs known to the system? The second part of the process, the generative part, is what we should call “imitation.” Here, the AI system is ostensibly imitating its own archetypal dog.

By doing so, however, Generative AI is not only automating imitation; it is also resurrecting terms and operations—and with them a very old way of discerning styles and of imitating images—which had long been inherent in the classical tradition of European art, and were canceled by late-modernist art theory in the second half of the 20th century. What modernist ideology had kicked out of the door is now coming back through the window of technology. Remarkably, even if clothed in stuffy Eurocentric terms, this new technology is global, and it works just the same with datasets of all sorts. The best demonstrations of style transfer one finds in technical literature these days use datasets of golden retrievers, not collections of Albrecht Dürer’s engravings. Whether we like it or not, every canon is a dataset; every dataset is a canon. This was always the case, conceptually, well before the latest wave of AI technologies made us aware of it by obliging us to work that way.

And whether we like it or not, the mode of functioning of today’s Generative AI is also reminding us that there is no generation without a dataset, hence no creation without imitation. As Renaissance scholars claimed, all creation inevitably involves some degree of imitation, and all imitation involves some degree of creativity. This is what early modern humanists called “creative imitation.” This is what formal analysis called reference to precedent: the awareness, acknowledgment, contestation, transgression, and criticism of precedent—of that canon of our own choosing which gives meaning to our voice. There is therefore a certain irony in that the latest avatar of electronic technologies may be reviving and vindicating traditional creative processes—some would even claim, universal creative processes—that modernist culture had safely, and we thought permanently, repudiated. Every Generative AI project today starts with the creation of a dataset. Is this not what the classical tradition—and, in more recent time, the pedagogy of formal analysis—always advocated? True, artificial intelligence is not particularly good (actually, is not good at all) when it comes to creatively contesting the datasets we feed into it. But are we any better? Artificial intelligence today obliges us to deal with precedent, but we no longer know how to do that. As imitation and style have been absent from most critical discourse in the visual arts for the past two generations, we are all, but particularly in the Eurocentric, modernist West, a bit out of practice and in need of retraining. We need to reengage with imitation and style in critical, interpretive, and creative terms. No artificial intelligence will do that in our stead—at least, not for some time to come.


[1] Peter Eisenman, The Formal Basis of Modern Architecture (Zurich: Lars Müller, 2006). Original Submitted in August 1963 for the Degree of Doctor of Philosophy, Trinity College, University of Cambridge

[2] Peter Eisenman, Ten Canonical Buildings, 1950–2000, ed. Ariane Lourie (New York: Rizzoli, 2008).

[3] Desiderius Erasmus, Dialogus cui titulus Ciceronianus sive de optimo genere dicendi (Basel: Froben, and Paris: Simon de Colines, 1528); Ciceronianus: or, a Dialogue on the Best Style of Speaking, transl. Izora Scott (New York: Columbia University Teachers College, 1908); also in Scott, Controversies Over the Imitation of Cicero (1910). In that dialogue Erasmus lampoons a fictional scholar called Nosoponus (“the sick guy”) obsessed with the imitation of Cicero’s Latin.

[4] See Alexander Nagel and Christopher W. Wood, Anachronic Renaissance (New York: Zone Books, 2010),150–57 and 289–300.

[5] See recent work by Mary Hvattum and in particular Style and Solitude (Cambridge, MA: MIT Press, 2023).

[6] See Harold Bloom’s The Anxiety of Influence, first published in 1973.

[7] See Leon A. Gatys, Alexander S. Ecker and Matthias Bethge, “A Neural Algorithm of Artistic Style,” ArXiv : 1508.06576 (August–September 2015); Leon A. Gatys, Alexander S. Ecker and Matthias Bethge, “Image Style Transfer Using Convolutional Neural Networks,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (Las Vegas, June 2016): 2414–23.



Log 58