Diffused History: Generative AI and the Crisis of Architectural Memory

Writing at the dawn of the social media era, the philosopher Bernard Stiegler considered the transformation of cultural memory under the digital conditions of a commercialised internet. Stiegler warned of the consequences of the industrialisation of “tertiary retentions,” the memory support objects (from photographs to videos) that help shape collective and personal identities. What happens, Stiegler asked, when this shared reservoir of cultural memory increasingly takes the form of a commodified online archive that we all uniformly draw from? He writes, “Being increasingly constituted by images and sounds that the media streams through my consciousness . . . my past is less and less differentiated from that of other people. It loses, therefore, its singularity, which is as much to say, I lose myself as singularity.”¹Bernard Stiegler, Symbolic Misery, Vol. 1: The Hyperindustrial Epoch, trans. Barnaby Norman (Polity Press, 2014), 6. For Stiegler this is not only a crisis of memory, but also a crisis of the future, since it is by drawing on the collective repository of tertiary retentions that we are able to imagine new forms of organisation and creativity. By standardising our available cultural archive, we also limit our ability to envision possible futures.

The rapid emergence of generative AI only amplifies these questions of digitally consolidated memory, the fragile status of collective history, and what remains of the creative potential of the future. Architecture, the focus of this essay, offers a particularly revealing lens through which to consider these developments, as a discipline perpetually negotiating between the inheritance of the past and forward-looking speculations. If social media turned cultural experience into a continuous stream of commodified images, generative AI transforms the entirety of that stream into a source of data for the production of enormous multi-modal models. Image generation tools like Midjourney or Stable Diffusion, for example, are trained on massive internet data scrapes, consolidated into datasets like LAION-5B (the 5B standing for 5 billion image-text pairings). These models are comprised of vast swathes of digitally recorded visual culture, including artistic production, architectural visualisations, amateur photography, advertising and graphic design, cinema and television. Although enormous in quantity, these datasets are, like the internet itself, also heavily biased towards Western and Anglophone content. This training material is embedded into latent spaces of compressed mathematical abstraction and available for endless extraction by diffusion models responding to our every prompt.

In the mathematical manifold of AI latent space, distinctions collapse between periods and mediums, high and low culture, realism and fantasy. Images with visual similarity are embedded in vectorial proximity and may influence in equal degrees generative outcomes.  Media theorist Roland Meyer describes generative AI’s tendency to reduce the particular characteristics of visual history to “combinable parameters within an extended category of ‘style.’”² Roland Meyer, “The New Value of the Archive: AI Image Generation and the Visual Economy of ‘Style,’” IMAGE: Zeitschrift für interdisziplinäre Bildwissenschaft 37, no. 1 (2023): 107. In Meyer’s formulation, everything, from Rembrandt to steam punk, becomes a style, a modifier or conditioning term available and transferrable to any object, scene or situation. In a lecture at the Max Planck Institute for Art History, he describes this flattening out of historical context as a generative AI mode of “generic pastness”; the models offer up generic images devoid of specificity, yet possessing an atmosphere of “pseudo-historical plausibility.”³Roland Meyer, “Generic Pastness: AI Image Synthesis and the Virtualization of the Archive,” research seminar, Bibliotheca Hertziana – Max Planck Institute for Art History, Rome, October 13, 2023, https://www.biblhertz.it/events/35511/2719701. A generated image of New York’s Guggenheim Museum, while clearly not accurate, possesses a certain Guggenheim-ness, or elements of Frank Lloyd Wright’s design “style,” according to Meyer.

This reduction of historical specificity to the generalities of approximate style is one facet of generative AI’s somewhat paradoxical relationship to the past. While AI models are indifferent to historical accuracy, they are nonetheless conservative in their orientation. Generative AI is clearly adept at reproducing statistical norms based on the preponderance of what already exists (online at least). What’s less certain is its ability to create something truly new or unusual. This predisposition has led Lev Manovich to assign the label “anti-avant-garde” to generative AI. He suggests that AI’s privileging of “familiar reality and its logic,” places the technology in opposition to the avant-garde impetus to break with tradition and disrupt normality. “Artists who want to use it for genuinely experimental work,” he claims, “are constantly fighting the medium’s own tendencies.”⁴Lev Manovich, A Medium That Thinks: AI and Media Cognition (2026), 9, https://manovich.net/index.php/projects/a-medium-that-thinks-ai-and-media-cognition (accessed May 29, 2026). And one of those tendencies is towards aesthetic conservatism, Manovich claims, going so far as to identify nineteenth-century realism and Paris salon-style academism as the dominant visual mode of AI image generators. “[C]lear, coherent, centered, and visually pleasing. The statistical dominance of this aesthetic biases models towards producing similar visual structures,” he explains.⁵Manovich, A Medium That Thinks, 10. AI’s horizon of formal novelty is therefore a kind of “bounded creativity,” Manovich claims, borrowing a concept from the philosopher of AI Margaret Boden.⁶Margaret A. Boden, The Creative Mind: Myths and Mechanisms, 2nd ed. (Routledge, 2005).

Turning now to the field of architecture, it’s somewhat surprising, given these conservative tendencies, that many of the architectural studios most associated with formal innovation have also been among the earliest adopters of generative AI. Zaha Hadid Architects, now led by Patrik Schumacher, and Thom Mayne’s Morphosis both claim to have integrated AI into their existing technological workflows. Mayne suggests that AI provides a means to integrate material complexity into projects in a manner beyond human capacity.⁷Ben Dreith, “Architects Need to ‘Take Responsibility for Shaping the World,’ Says Thom Mayne,” Dezeen, February 1, 2023, https://www.dezeen.com/2023/02/01/thom-mayne-aia-talk-new-york/. By 2023 Schumacher was already speaking of the use of text to image generators like DALL-E 2 and Midjourney within the Zaha Hadid studio. He stated at that time, “Not every single project is using it, but let’s say most – I’m encouraging everybody who’s working on competitions and early ideation to see what comes up and just to have a larger repertoire.”⁸Nat Barker, “Zaha Hadid Architects Developing ‘Most’ Projects Using AI Images, Says Patrik Schumacher,” Dezeen, April 26, 2023, https://www.dezeen.com/2023/04/26/zaha-hadid-architects-patrik-schumacher-ai-dalle-midjourney/. In an odd instance of recursivity, the descriptor “Zaha Hadid” is frequently used in the process of prompting. The studio’s trademark style of fluid, curvilinear forms (pioneered by its founder and subsequently theorised by Schumacher as “parametric design” due its algorithmic, parameter-based approach) is extensively represented online and therefore deeply embedded within the latent spaces of contemporary generative models. In this strange feedback loop, Hadid’s digital parametricism, once considered a radically future-oriented design language, returns through generative AI as a statistically reproducible style category: no longer an avant-garde rupture with architectural history, but yet another aesthetic atmosphere circulating within the generic pastness of the model. Schumacher claims the process is not significantly different than previous techniques of idea generation that referenced prior projects. He insists, “I accept all of that into our oeuvre. Any of what comes out of this, I claim authorship for it in terms of validating, selecting, elaborating. So I feel very kind of empowered by all this possibility.”⁹Barker, “Zaha Hadid Architects Developing ‘Most’ Projects Using AI Images.”

Architectural designer Tim Fu began his AI experimentation as a member of Zaha Hadid Architects’ computational research group before branching out to form his own Tim Fu Studio, self-described as “an architectural studio organised to integrate AI into core practice.”  Forging a style that fuses the parametricism of his former studio with the inherent aesthetic conservatism of generative models, Fu advocates for a form of AI-automated “neo-classical futurism.” His generated images of contemporary interiors and exteriors incorporate ornate Corinthian columns and almost biomorphic, gothic and baroque structural features (see Figure 1). It’s a technologically enabled maximalism that eschews the minimalism and functionality of the often-venerated modernist architectural avant-garde. Fu asks, “So why not bring back ornamentation, bring back the beauty and the aesthetics that we once held so highly in the classical era, and also allow machines to continue to fabricate and produce feasibly for us and free us up to do the more intricate and beautiful parts?”¹⁰Hannah Carlsson, “AI Design Could ‘Bring Back the Beauty and Aesthetics of the Classical Era,’ Says Tim Fu,” Dezeen, August 16, 2023, https://www.dezeen.com/2023/08/16/ai-aesthetics-classical-era-tim-fu-aitopia/. Yet the architectural horizon implied by these images remains surprisingly narrow: generic and familiar features from architecture’s distant past fused with equally familiar elements of recent futurism. And can the forms, aesthetics and typologies of architectural history be excavated and repurposed without also bringing forward at least a remnant of the cultural associations and hierarchies of their original context? Despite their attempts at formal novelty, Fu’s AI-generated images obey the visual conventions (polished surfaces, axial symmetry, dramatic, but harmonious lighting) of contemporary luxury architectural render culture. It’s perhaps telling that the studio’s first (AI-led) commission is a rather conventional brief for a luxury residential complex of six high-end “villas” on the shore of Slovenia’s Lake Bled.

There is, however, a more troubling recent convergence of AI and architectural conservatism and a more disturbing vision of neo-classical futurism. As the cultural critic A.V. Marraccini has identified, a techno-libertarian project has recently emerged with a particular affinity for AI-generated classical architecture. Praxis, a company established in 2019 by two founders in their twenties, Dryden Brown and Charlie Callinan, now describes itself as an “online community starting a new city” and the “world’s first digital nation,” with the stated purpose to “revitalize Western Civilization by creating a new culture focused on the pursuit of heroism, truth, and beauty”¹¹Praxis Nation, archived May 21, 2026, Internet Archive, https://web.archive.org/web/20260521203230/https://www.praxisnation.com/. (see also Figure 2).  Currently boasting an online “population” of 2,653, Praxis has the much larger ambition of creating an exclusive 10,000-inhabitant city in the Mediterranean committed to transcending the current limits of human civilisation through a dedication to beauty, nobility and the embrace of advanced technologies. Brown and Callinan claimed to have already raised half a billion dollars towards this seemingly unlikely goal from Silicon Valley investors like the Peter Thiel backed city development venture capital firm Pronomos Capital and the cryptocurrency investment company Paradigm Operations. The object of some degree of journalistic fascination, Praxis has been variously described as a “tech-bro utopia”¹² Joseph Bullmore, “Praxis and the Promise of a Tech Bro Utopia,” Spear’s, February 8, 2024, https://spearswms.com/wealth/column/praxis-promise-tech-bro-utopia/. and an Ayn Rand-inspired “crypto-city.”¹³Joseph Bernstein, “Who Would Give This Guy Millions to Build His Own Utopia?,” New York Times, December 12, 2023, https://www.nytimes.com/2023/12/12/style/praxis-city-dryden-brown.html.

Praxis officially announced its city building aspirations in 2023, complete with digital renders of the urban plan produced by none other than Zaha Hadid Architects (see Figure 3), who were at that time touted as the designers overseeing the project (although the project is conspicuously absent from the studio’s current website).

While the Zaha Hadid renders of the proposed tech-utopia are an amalgam of contemporary minimalism and neo-classical axial planning, the Praxis website articulates an even more imperial reference point: “Praxis is the first Digital Nation, dedicated to a monumental idea, one that has been suppressed for the better part of a century: Rome. Rome is the symbol of the people of the West, and carries within it a concept of heroism as a path to the transcendent.”¹⁴“Life After Labor: AI Will Return Us to Tradition,” Praxis Nation, archived May 21, 2026, Internet Archive, https://web.archive.org/web/20260521201338/https://www.praxisnation.com/content/life-after-labor. Praxis and its founders also have a penchant for posting AI-generated images of the neo-classical variety on their social media accounts, illustrative cityscapes complete with marble statuary and pillared rotundas (see Figure 4). These images seem to embody the movement’s techno-reactionary ethos, epitomised by Praxis slogans like, “AI will return us to tradition.”¹⁵“Life After Labor” In her poetic, yet trenchant critique of the Praxis project, Marraccini questions its ambitions to transcend the conditions of our world in favour of a fantasy image of idealised pastness, insisting, “There is no world without historical constraint.”¹⁶A.V. Marraccini, “The Mutilated World,” Verso Blog, April 12, 2023, https://www.versobooks.com/blogs/news/the-mutilated-world. Italics in original. Voicing her objections from the position of a queer, “broken-bodied” person who would never be granted citizenship within the Praxis utopia, she explains: “Exiles from real and imagined city-states, after all, are often particularly attendant to history, and particularly aware of the pitfalls of claiming to remake, renew, or erase it, whether rhetorically or with images alone.”¹⁷A.V. Marraccini, “The Mutilated World,”

While the techno-reactionary Praxis project embraces the conservative tendencies of AI-generated architecture, other architects, artists and designers (often operating at the fringes of their disciplines) are challenging the technology to envision alternative urban pasts and futures.  We could point to the uncertain and provisional images of Italian architect Cesare Battelli, full of scaffolding and fibrous materials, as if in a perpetual state of construction or even biological growth. Battelli’s generated images tap into the more eccentric traditions of architectural drawing – from Piranesi’s etchings of impossible prisons to the chaotic geometries of Lebbeus Woods – producing images that are somewhat familiar, yet also strange and disorienting. Or we could explore the generated cityscapes of Egyptian-American architect Hasan Ragab that disrupt the Western, symmetrical and sanitised biases of AI platforms like Midjourney or Stable Diffusion. Ragab compels these models to conjure up speculative urban forms based on the ancient and contemporary typologies of Alexandria and Cairo, rather than those of Rome or Los Angeles.  His “The City as a Tram” series, for example, imagines the humble Alexandrian tram as the foundational recurrent building structure of a sprawling, mobile and adaptable Arabic urbanism. Or finally, we could turn to the “Auntieverse” project developed by the Singaporean designer Wenhui Lim under the name “niceaunties,” an imagined alternative or future world where the often-marginalised figure of the Asian “auntie” is the focal point for an urban landscape built in her image. Lim generates animated and still scenes of aunties (often taking on gigantic proportions) running factories, devouring sushi, performing surgery, and enjoying beauty treatments. These works confront the invisibility of the elderly in both society and in AI model outputs, presenting a universe that celebrates, rather than neglects, older women. All of these designers find ways of pushing AI generation beyond a condition of “generic pastness,” managing to provoke AI image archives into new and unexpected configurations.  Rather than treating AI as a reservoir of dominant, established forms, these artists and architects use the technology to recover and mobilise suppressed, marginalised and over-looked elements of visual culture. Most of these images do not translate seamlessly into material for conventional briefs or commissions, but they do help expand our sense of architectural possibility beyond the familiarity of existing repertories and conventions. This is not to say, however, that even the more innovative uses of generative AI are immune to commodification. Lim’s colourful and playful Auntieverse has proven appealing for collaborations with brands like Swatch and L’Oréal. The distinctiveness of these images and their ability to overcome the generic aesthetics of generative models can give them commercial, as well as cultural value.

When Marraccini advises us that there is no world without historical constraint, she is also offering a reminder that history is much more than simply a repository of past forms or a latent archive of available styles ready for extraction or diffusion. History is a site of conflict and contestation that conditions the range of futures available to us. While the default mode of generative models is one of an aesthetic conservatism that synchronises well with the political and financial conservatism of both luxury developments and techno-utopias, it remains possible at least to push AI technologies in different and more promising directions. Our architectural future may depend on it.

Joel McKim

Joel McKim is Reader in Digital Media and Culture, Director of the Vasari Research Centre for Art and Technology and co-director of the Centre for Creative AI at Birkbeck, University of London. He is the author of Architecture, Media and Memory: Facing Complexity in Post-9/11 New York (Bloomsbury) and is currently co-editing a special issue of the journal Theory, Culture and Society considering forms of morphological and computational intelligence entitled “Patterns: Between the Discrete and the Continuous.”