Allies in Exile
In a chapter of Diffusions in Architecture - Artificial Intelligence and Image Generators, edited by Matias del Campo, I argue that in the context of prompt-based architectural visualization, designers must learn to navigate landscapes beyond the technical.
Read the full paper here.
Michael Chabon's 2007 novel "The Yiddish Policemen's Union" is set in an alternate history in which the US government provisioned land in Alaska in 1940 for the refugee settlement of European Jews fleeing Nazi persecution. With the fledgling State of Israel destroyed in 1948, European Jewry found a new home not in New York, but in Sitka - a large, Yiddish-speaking metropolis initially established as a small temporary settlement on unceded Tlingit tribal land. The plot centers on the illicit construction of an architecture, and unfolds within atmospheric world that blends Jewish and Tlingit culture in the Alaskan landscape. Thematically, this work that is set in an alternative present explores a breadth of issues that resonate with our actual present, including: identity and community; extremism and assimilation; and cross-cultural contact, friction, and appropriation. Chabon’s writing is vivid, and since reading this work I’ve been a bit obsessed with his imaginary Sitka - something like a 19th century Vienna in the Arctic. What would the material culture of this place be like? How would the trauma and traditions of people from the "old world" experience modernity in this unique climate and cultural situation? Would Jewish design and culture harmoniously meld with the indigenous traditions (Tlingit talit, anyone?), or would this alternate history rhyme with our own, with the transplanted heart of Jewish life beset by issues of exploitation, appropriation, and oppression?
I raise Chabon's opus of speculative fiction in this context for two reasons. Foremost, as you can likely tell, I'm a fan. This is evidenced by the nearby synthetic images - speculations set in Chabon's alternative present. These explorations “worldbuild” the architecture and artifacts of a place that, had history gone a slightly different direction, I might have been an citizen of. But beyond these largely personal indulgences in fan-fiction, I would argue that Chabon's novel opens a useful conversation for all of us who operate at the intersection of generative AI and architecture.
Working with image-synthesizing CLIP-guided diffusion probabilistic models, or "diffusion models", feels different than CAD. It even feels different than other forms of architectural image-making. In contrast with traditional visualization, working in this way feels less like constructing visions of the future, and more akin to conjuring alternate versions of our present. Perhaps this is because the dazzling synthetic images that draw so much attention on Instagram are not de novo visions of the future drawn from the mind of a sole creative author. Rather, these are inferences drawn from data scraped from our collective past - data laden with the visual richness of, as well as the latent biases of, the cultures from which they were extracted. In this way, we are all engaging in a form of "fan fiction" at the moment we invoke in our prompts the proper name of an author, or an architectural style, or mention a specific geographical region or culture that manifests that unnameable quality that we seek. It is undeniable in this moment that we're working with a technology of culture, one that bears the indelible mark of the context that produced it. In stark contrast with traditional CAD, which assumes a posture of objectivity and neutrality, diffusion models insist that we address many of the same issues raised by Chabon's work - issues of identity, community, extremism, appropriation, and exploitation.
In my early engagement with these systems, I felt compelled to address the issues mentioned above, but also felt hopelessly under-prepared to face them. With nothing in my education as a computational designer to guide me, I found myself instead turning to my personal experiences - to memories, stories, and cultural references that I knew well. The nearby images, for example, show scenes and forms drawn from my own childhood - nonce orders assembled from palmetto grass, temporary structures assembled from the wreckage of the Challenger disaster, houseboats surrounded by manatees that can no longer survive in the wetlands of Northern Florida.
In many ways, all of us in the design computation community are under-prepared to face the issues raised by this new digital "turn" - a term we use to describe moments of disciplinary crisis. The imagery that has dominated architectural social media across the past year was produced through methods that are unlike anything we've dealt with before. It seems we can forget everything we know about parametric data trees, object-oriented programming, and digital morphogenesis - this is fan-fiction all the way down. Such a shift holds narrow ramifications for those of us who work in design computation, and also broad implications for the discipline and the position of technology within it.
Those of in design computation have a daunting task in front of us. As stewards of the cultural and social dynamics of the discipline's engagement with technology, it falls to us to provide a critical assessment of software tools employed by architects, a record of emerging methods of design enabled by these tools, and an accounting of how the affordances of these tools and methods intersect with creative practice. Our role takes on particular significance in periods of pronounced technological change, when the dynamics of tools and practice are seen in stark relief. In many respects, we've been here before. Many of us in this community served in such a capacity during the last digital "turn", when parametric and generative design entered into the architectural consciousness in the early 2000s.
This time is different, and we're not prepared.
The last time, our community called upon a specific constellation of allies to cope with disciplinary questions we couldn't adequately address on our own. In this earlier period, a specific set of organizations, events, and collectives were established (see, for example the Smartgeometry group founded by Robert Aish and others in 2001, or the Institute of Computational Design at the University of Stuttgart), and served to bring architectural practitioners and academics together with a range of expertise not typically required by our discipline. Given the technologies at play, this included some figures that we might expect - such as computer scientists, structural engineers, and professional programmers - as well as some we might not - such as systems theorists, computational geneticists, and artificial life enthusiasts. Such an eclectic constellation of voices were required in that moment in order to further the broader project, and address critical questions regarding the integration of a specific set of emerging and extrinsic design methods (e.g. optimization, emergence, morphogenisis) into architectural practice.
This time - both in terms of the nature of the technology, and the nature of the surrounding questions - things are different. In some of the images shown nearby, I've sought to address some of the ways that issues of identity, appropriation, and culture are unavoidable when working with diffusion models. For example, since language itself is a cultural signifier, the series of images on the far left illustrates the unavoidable non-neutrality of text-to-image systems. Consider that the software that produced this set of images, each invoking the same prompt translated into four different languages, would operate differently for different users - perform differently in Kansas than it would in Korea or Columbia. Extending this small example to computer-augmented design more broadly, the implications are sobering. While our tools have never been truly neutral, how might design practice adapt to working with systems so deeply intertwined with culture, and that can no longer mistaken as neutral?
As in the previous turn, this new technology suggests a set of issues extrinsic to our discipline. Just as before, we will need allies, but the specific constellation will certainly be different. Some of these voices we require will be technical - including data scientists and machine learning engineers - but the most critical guidance may come from elsewhere - such as data ethicists, crypto artists, historians, and anthropologists. Two subject areas hold particular relevance, and are worth expanding upon - each emphasize the cultural role of computation, and have historically considered “external to the discipline” of architecture. First, software studies is an interdisciplinary research field that accounts for the social and cultural effects of software systems, and includes such figures as Benjamin Bratton and Lev Manovich (a contributor to this text). While a better connection with this community is long overdue in design, an approach that understands CAD through the lens of media - that is, as an instrument that connects us to knowledge, skills, and abilities that we would not otherwise hold - would prove particularly useful in this moment. Further, figures versed in digital labor studies, such as Trebor Scholz and Niloufar Salehi, would help us unpack the techno-social power dynamics at play in this new landscape - including issues of exploitation and appropriation surrounding datasets. In many ways, the introduction of text-to-image into architectural visualization work represents a classic technological labor disruption of the sort that produces clear socioeconomic "winners" and "losers". While typically, the winners in such a scenario tend to be the tech-savvy, the dynamics here may be different, in that the lowering of barriers has already flooded our market with compelling and imaginative images, each of which was made in seconds. As any labor economist will tell us, when a product becomes much easier to produce, this project is devalued. It would seem that we're entering a phase in which architectural image-making is easy, cheap, quick, and potentially holding radically less value than it has held in the past. How will the flood of incredibly beautiful and radically cheap images change our discipline?
Like Chabon's refugees exiled in Sitka, there are those of us in design that may feel suddenly thrust into a foreign landscape, daunted by the technical complexities of a new technological domain, and unprepared for the ethical implications thereof. It falls to those of us who advocate for the humanistic applications of technology, together with the allies mentioned above and others, to assist practitioners in critically and ethically engaging with this powerful new computational "tool of the imagination", and to help foster positive cultures of practice around it. A central question remains: to what extent will ML tools catalyze a shift in the centers of power in design practice? Along with clients and contractors, developers and regulatory bodies, architects are but one small part of a broader network of social, technical, financial, and cultural exchange. In a context in which architectural labor is already beset by inequities and exploitation, how might the seemingly-inevitable changes on the horizon exacerbate these existing problems?
Or, worse yet, might things not change at all? Might we miss a rare opportunity for transformation, and fail to bring the discipline more in line with our values?