AI Productivity and Creativity
Patrik Schumacher
Published in: Williams, A.(ed), Five Critical Essays on AI, TRG Publishing, London 2025
There is a lot of ambivalence in architecture, as well as in society at large, about the effects of Artificial Intelligence. For me and my firm the advent of AI has been an unmitigated boon. We do not see or feel any loss of agency. Quite the opposite: we feel empowered. We do not feel threatened by the proliferation of epigones and competitors. Quite the opposite: we are confidently exploiting the new tools to pull ahead of our competition. I also do not fear that fees will be squeezed. Previous productivity boost – like the introduction of CAD or later BIM – did not disrupt the historical trend increasing professional fees as percentage of construction. These productivity gains went into service and product quality upgrades rather than the reduction of employment.
At Zaha Hadid Architects the adoption of AI image generation systems like DALL-E, Midjourney and Stable Diffusion a.o. was rapid and pervasive. The use of these tools – together with our own tool building efforts on top – has delivered an important capacity and productivity boost, especially for early ideation and visualisation. While most front-end designers are using various generic and bespoke tools, we have also a dedicated AI development team within our computational design group CODE.
We have developed plug-ins for both Maya and Rhino that allow our designers to instantly render models in AI. We are mostly building on top of Stable Diffusion and are training specific LoRA (Low-Rank Adaptation) models from our own image bank for every new project to better target and streamline the AI ideation. We have also started to use a mediating platform – xfigura.ai – that makes many underlying AI models available and allows the AI project workflow to be displayed on a canvas that can be shared within the team and that allows the workflow to be graphically displayed with nodes and links in analogy to grasshopper. We are also using various AI video generation tools to upgrade our presentations with animations.
In the meantime, some start-up companies have emerged that offer AI tools specifically to architects: LookX.AI is a Chinese generative image AI platform created by XKool, who had been using AI tools for years to optimize massing models within planning envelopes. LookX supports the training of models by users as well as working in teams. Gendo AI is a UK based platform crafted specifically for highly plausible, photorealistic architectural visualisations, focussing on ease of use and quick uptake within bread and butter architectural and interior work. The industry of is moving fast, so this short snapshot provided in this article will be obsolete very soon, if not already at the time of printing.
In any event, I can say with conviction that all these new AI image tools already deliver not only significant productivity gains but indeed a notable creativity boost. Ever since DALL-E 2 prompting results were credible and coherent with respect to perspective, light & shadow, poise etc. while at the same time delivering rather novel compositions. The ability to prompt with references to various architects – mostly ourselves – mixed with other architects (like Frei Otto), other styles, non-architectural industrial products or structures, landscape features etc. shows generative image AI to be an engine of invention via hybridization. The images are coherent and suggestive enough to allow for their re-creation via 3D modelling to bring them into the designer’s proper medium. The speed, number and variety of options with which ideas can be explored implies a much enlarged search space to find spatio-morphological ideas that can then be worked up into sketch solutions, or first approaches towards a potential design solution. The photorealism of the ideation images implies the ability to bring clients early on into the ideation process. This can be very efficient. The capacity to sift through higher numbers of options, especially a larger range of options, implies an expanded and more thorough search which in turn implies an increased level of design process rationality and decision making. A solution that is selected from 250 options or candidates, from within a larger range of possible solutions, should inspire more trust than a solution that was selected from 5 or 10 candidates. To me it is evident that an design ideation process that is augmented by generative image AI is both more creative, allowing for more novelty, which in turn gives more chances to discover innovations (novelties that represent improvements).
While I am thus enthusiastic about this new wave of powerful AI tools, I do not think that these generative image tools advance the discipline to the same extent as the computational-algorithmic design tools that started to enter the avant-garde arenas of our discipline since the mid-1990s and that led to the formation of a whole new style: parametricism. This style was responding to the challenges posed to urban and architectural design by the socio-economic restructuring from Fordism to Post-fordism triggered by the micro-electronic revolution. The response involved the belated introduction of the very computational technologies within architecture which had already transformed the tasks and problem space of architecture. This combination of the historically new design tasks – new level of complexity and dynamism in urban life processes – with radically new design tools bringing on radically new concepts and forms, ushered in this new style which evolved and matured in symbiosis with the computational tools over the last 30 years. The latest stage of this development within architecture is the subsidiary style of tectonism, to be understood as parametricism’s most sophisticated incarnation. The recent proliferation of AI image tools does not imply yet another style, but it does indeed imply a huge potential for the accelerated adoption of parametricism, in particular by the youngest architects. While the built environment does not yet manifest this, it is evident in the world of architectural AI images that circulate on the internet and manifest the desire and aspirations of young architects and designers. If you google AI architecture, you can immediately confirm this. At least 80% of AI generated architectural output can be classified as parametricism. For me this is deeply satisfying, in contrast to the stale and retrograde output of the profession for actual construction. The simultaneous advent of the metaverse is starting to engage in a combustive synergy relation with AI. The in my view inevitable prospect of the metaverse taking over large parts of the internet and substituting or augmenting large parts of the physically built environment implies an avalanche of design work that can only be coped with by the full utilisation of AI empowerment. Here too, with respect to metaverse design, we can already unmistakeably see the coming hegemony or all-pervasiveness of parametricism.
While the big AI models know all about Zaha Hadid Architects and parametricism and are virtuoso contributors to parametricism – after all, we and many other protagonists of parametricism have been feeding images into the internet for over 25 years – the same is not yet true for the more recent and more exacting style of tectonism. Here the relative rarity of pertinent example images feeding into the training of the big models means that convincing results cannot yet be generated. There is also the added difficulty posed by the much higher level of specific rationality conditions that tectonism represents. Tectonism relies on various computational engineering and optimisation tools, integrating structural engineering logics, environmental engineering and optimisation for computational fabrication, i.e. there are many ways in which the AI model can get this wrong. However, the case is far from hopeless. We have already been teaching our AI systems the various engineering logics and rules of tectonism with convincing success by using our extensive image bank, plus simulated synthetic data, for training purposes. There is, however, another potent, and indeed more important contribution AI can make to the further advancement of tectonism (and thereby of architecture): The Large Language Models (LLM) like ChatGPT are helping us architects (who are not professional software developers) to programme many more of the computational design tools we would like to have to advance the rationality, functional sophistication and intricacy of our designs. At Zaha Hadid Architects about 50 of our 500 members of staff are engaged in research, and for us research usually materialises in computational design tools that upgrade our design intelligence and workflows. In the medium and longer term this aspect of the power of AI is probably significantly more impactful than the image generation tools. The image models themselves too can be engaged in this more functionally driven tooling. These models can be trained with simulation results, for instance of structural topology optimisation or of heatmap outputs of agent-based occupancy simulations, and then deliver approximate, usable results in a fraction of the time required by the original simulations, thereby delivering very useful shortcuts for the designer churning options with nearly real-time analytical feedback.
To conclude: I am not only optimistic, I am thrilled about AI empowering architecture and literally all other arenas of work, including science as well as creative work. However, it remains the task of self-reflective, mission driven humans – interacting with each other in critical discourses – to device, train, steer, put to work and utilize AI systems to the empowerment and flourishing of the human project.