A city is never simply what we see in the present. It is the product of successive transformations, substitutions, adaptations, and continuities. Each era adds something to the one before it, altering the way it is read without necessarily erasing what came first. In this sense, the city can be understood as a palimpsest: a structure in which different periods remain present and legible through architecture. It is an idea first articulated theoretically by urban historian André Corboz, who described the territory itself as a text that is continuously rewritten, yet never entirely erased.
The contemporary challenge is not simply to preserve what already exists, but to understand how a new layer of time can be added without severing continuity with those that preceded it. There is no single way to engage with the existing fabric. Historic architecture may be preserved, reconstructed, reinterpreted—or confronted with a deliberately contemporary language.
Berlin’s Neues Museum, designed by David Chipperfield Architects between 1997 and 2009, offers one possible approach. Surviving elements were consolidated, while lost portions were reconstructed through contemporary materials and forms. The new remains clearly identifiable without feeling alien; the building’s renewed spatial continuity never erases the traces of its history.
Herzog & de Meuron pursued a different strategy at Am Tacheles, also in Berlin. The project, developed between 2014 and 2019 and built from 2018 to 2024, begins with an understanding of the city as a product of layers, demolitions, and reconstructions. Its relationship with history does not emerge through the replication of vanished architecture, but through the structure of the urban block, the preservation of the existing fragment, and the construction of an unapologetically contemporary new form.
Another possibility is to reject any attempt at stylistic continuity altogether, making the distinction between two periods explicit. Berlin’s Fire and Police Station, designed by Sauerbruch Hutton and completed in 2004, stands alongside a historic 19th-century structure. Here, the new architecture makes no attempt to blend into what came before. The brick volume of the original building is juxtaposed with a new structure wrapped in a highly distinctive skin of colored glass. The contrast in material and color makes the arrival of a new era unmistakable.
Yet before any of these architectural strategies comes another moment: the stage at which a design proposal is neither continuity nor contrast, but simply an image to be questioned. It is within this territory—before a project becomes structure, envelope, and construction site—that generative artificial intelligence may begin to play a compelling rolein the architecture of heritage.
Intervening in a historic context requires more than familiarity with a particular style. It demands an understanding of the underlying structure and rules that define a specific place.
In 2022, the study Automatic generation of architecture facade for historical urban renovation using generative adversarial network explored this possibility through a generative network trained to produce new façade configurations from a dataset of historic buildings along Harbin Central Street, in northern China. The aim was not reconstruction, but rather the creation of a tool capable of abstracting recurring characteristics of the built heritage and using them to generate alternatives.
It marks an important conceptual shift: heritage is treated not merely as something to be preserved, but as a source of architectural knowledge.
By 2025, this line of research had moved toward diffusion models. The study Preserving architectural heritage in urban renewal employs Stable Diffusion alongside LoRA—Low-Rank Adaptation, a lightweight model used to define a specific expressive identity—and ControlNet, a neural framework that constrains the geometric consistency of generated results. Together, these tools are used to produce historic-style façades from targeted datasets and compositional references.
The authors themselves, however, acknowledge that substantial limitations remain: the semantic complexity of architecture, the need for highly specific training, and the continuing importance of human oversight.
Despite the growing body of research on AI applied to architectural heritage, real-world applications of generative AI that result in the construction of new buildings within historic settings remain exceptionally rare. Recent literature does not yet describe generative AI as a technology fully integrated into conventional heritage design and construction workflows. Most research continues to focus on image generation, digital reconstruction, simulation, or decision support.
A 2026 review published in Building and Environment confirms this systematically: images remain the dominant form of data, both as input and output, while systems capable of integrating generative AI with geometry, BIM, and performance-based workflows are still comparatively underdeveloped.
The reason is not simply caution toward a new technology. Generative AI emerged primarily as a tool for representation and visual synthesis, while architecture must also resolve structure, feasibility, and performance. Within the field of historic preservation, further layers of complexity arise: authenticity, reversibility, material compatibility, and responsibility for the intervention itself. In this context, no image—however convincing—can resolve these questions on its own.
For now, then, the most tangible contribution of AI to architectural heritage is not the completed building, but the possibility of anticipating, visualizing, and debating different configurations before they become physical matter. This is not a question of mistrusting the technology, but of recognizing the current maturity of the process.
Cities, in any case, will continue to change. The challenge is not to prevent transformation, but to guide it.
The Neues Museum and the Fire and Police Station demonstrate that the relationship between new architecture and the existing city can take radically different—and equally legitimate—forms. Artificial intelligence does not introduce a third architectural position. Instead, it makes the complexity that precedes any design decision easier to explore.
If the past is regarded solely as something to be preserved, design risks turning the historic city into a static image. But if it is understood as a network of relationships and accumulated knowledge, it can instead become one of the conditions through which the future is imagined.
The city as palimpsest is therefore more than a metaphor for memory: it is a design condition. Every new building adds another layer—one that will, in time, become part of history itself.
Generative artificial intelligence can contribute to this process not because it knows which future ought to be built, but because it allows us to visualize it as a hypothesis: to compare possibilities that do not yet exist, and to question the relationship between what has been and what might yet be.
The future of the historic city need not be forced to choose between memory and contemporaneity. Imagining what does not yet exist means giving architecture the opportunity to question itself through forms we may not have seen before.
Alessandro Peritore