Generative AI is becoming increasingly visible in architectural practice, but much of its current use still sits near the end of the design process. The architect develops the project, then AI is brought in to sharpen the image, change the atmosphere, test materials, or produce a more convincing presentation. That can be useful, but it places generative capacity at a moment when many of the important architectural decisions have already been made.

The more consequential opportunity will be to introduce AI earlier, in the middle of the design process, while the architecture is still open to change.

At the massing and schematic stages, a building is still a hypothesis. Height, volume, program, structure, circulation, and environmental response remain negotiable. This is where variation can influence the architecture itself rather than only its representation.

From Recreation to Process

This work began through a collaboration with the Gaudí Foundation to explore whether generative AI could recreate architectural qualities associated with Gaudí. The initial question was largely one of reproduction: could an AI system learn from Gaudí’s work well enough to generate convincing new architectural propositions?

Studying and reflecting on how Gaudí actually designed made clear that reproduction was not the most consequential question. His work suggested a different direction: rather than asking how AI might reproduce a recognizable formal language, we began asking how it might participate in a design process based on conditions, response, iteration, and judgment.

The focus therefore shifted from recreating appearance to understanding how generative systems might enter earlier in design, while form was still being established, challenged, and revised.

We refer to the resulting approach as In-Process AI: generative models enter while the architecture is still unresolved rather than primarily after the design has been completed.

The architect begins with an authored three-dimensional massing. That 3D massing is read by the AI system and supplemented through a contract that establishes what each region is, how large it is, what should remain fixed, and what may change. The AI system then develops architectural propositions in an intentionally unfinished line-drawing representation. The architect selects a direction, and the chosen proposition is returned to editable geometry so the design process can continue.

Visualization remains part of the workflow, but it sits outside this generative design loop. Material, lighting, and presentation can follow later; they do not replace the return to geometry. The generated image is treated as an intermediate design proposition rather than as a final product.

A Precedent in Process, Not Style

This shift from appearance to process made Gaudí’s working methods more relevant than his formal language.

His hanging-chain models, developed most fully for the church of the Colònia Güell, worked by establishing conditions and allowing another system to answer back (Tomlow 1989). Loads were applied, the chains found equilibrium, and a change to one part altered the whole. The form was not selected from a set of finished images; it emerged through an iterative relationship between designer and system.

This belongs to a broader lineage of form-finding. Frei Otto similarly used physical processes to generate form rather than imitate its appearance (Otto and Rasch 1995). Gordon Pask described design as a conversation rather than a command (Pask 1969), a useful distinction here: a prompt can be transactional, while a design process depends on repeated response, adjustment, and judgment over time.

Donald Schön described the same pattern from inside practice as a “reflective conversation with the situation”: the designer makes a move, the situation talks back, and the designer judges what to do next (Schön 1983).

The comparison is useful, but only to a point. Gaudí’s hanging models were governed by physical forces, while generative image models work through statistical probability. A chain settles into equilibrium because gravity requires it to; an image model can produce a form that appears structurally resolved without having tested whether it is.

The precedent is therefore not Gaudí’s style, nor the physical certainty of his models. It is the idea of designing by establishing conditions, receiving a response, judging that response, and adjusting the process accordingly.

Keeping Judgment in the Process

That distinction is why the architect remains central to the workflow. The architect decides what is fixed before generation begins, adds the information that guides elaboration, evaluates the alternatives, and decides which proposition moves forward. Selection happens while the representation remains intentionally unresolved, before later visualization can make a proposition appear more complete than it is.

This is not presented as a solved system. At the moment of selection, structural validity, daylight performance, cost, buildability, and reconstructability have not yet been fully tested. The architect is making a consequential decision with incomplete information, as is also common in conventional early-stage design, but here some of that uncertainty is intensified by the gap between generated representation and measurable geometry.

The focus is therefore on the points at which judgment enters the workflow, and on what information is actually available at each of those moments. The central question is: Where in an early-stage design pipeline can human judgment be inserted, and what can the architect actually assess at each of those points?

The In-Process AI Method

The method below puts that focus into operation. It is divided into three parts: a phase before the loop that establishes the starting condition, the In-Process AI loop itself, and a later visualization phase after the loop. Within the loop, control passes deliberately between architect and AI system through interpretation, instruction, generation, selection, reconstruction, and evaluation.

Although the workflow emerged from the Gaudí Foundation collaboration, the method is illustrated here through a generic residential massing to test whether the process can operate independently of Gaudí’s formal language.

Before the Loop: Establishing the Starting Condition

Before In-Process AI begins, the architect establishes the initial position of the project and translates it into an authored three-dimensional massing. This stage is not simply about producing geometry. It is where the architect determines the central spatial idea, identifies the priorities that should guide later development, and begins to define what the project should and should not become.

AI may already be used conversationally during this phase to test assumptions, raise alternatives, or expose potential weaknesses in the initial idea. Its role here is closer to sparring than generation: helping the architect sharpen the proposition before the 3D massing becomes the spatial input for the workflow. As generating and developing alternatives becomes increasingly fast, this early act of defining a strong starting position becomes more important rather than less.

This phase ends with the authored massing. This 3D model forms the first fixed spatial condition of the workflow and becomes the point at which the loop begins.

The In-Process AI Loop

1. Read the 3D Model

The loop begins when the architect-authored massing is provided to the AI system. At this point, the project shifts from a primarily conceptual proposition into an explicit spatial input. The AI system is asked to interpret the geometry, scale, organization, and relationships already established by the architect rather than generating a project from a text prompt alone.

This first reading creates the base from which all later transformations are made. The existing 3D massing therefore remains visible as a reference condition throughout the process, allowing later propositions to be understood as transformations of an authored starting point rather than independent images.

2. Prime the Spatial Information

The geometric input is supplemented with information that makes its spatial logic more legible to the AI system. Dimensions establish scale, colour coding differentiates parts of the massing, and a clear layer structure organizes geometry into groups that can be referred to consistently.

This dimensional, visual, and organizational priming reduces the amount of spatial information that has to be communicated through language alone. Rather than asking the AI system to infer the meaning of an undifferentiated mass, the architect creates a readable framework in which geometry, dimensions, layers, and labelled regions can be referenced in the following steps.

3. Define the Contract

The architect then develops the design contract through a structured process of questioning and clarification. Rather than relying on a single prompt, the AI system can be instructed to continue asking questions until ambiguities around the architect’s intentions, priorities, constraints, and areas of freedom have been sufficiently resolved.

The resulting contract establishes region identities, fixed versus free elements, spatial priorities, and annotations linked directly to the geometry. Together, these define how the primed 3D model should be interpreted and transformed.

The purpose of the contract is not to describe a finished building. Over-specification would remove much of the value of generation. Instead, the architect establishes a boundary between constraint and freedom: enough information to preserve the central architectural idea while leaving parts of the proposition unresolved.

Annotations connect instructions directly to spatial conditions in the 3D model. Information can be attached to specific parts of the project, so different regions carry different constraints, intentions, and degrees of freedom.

4. Develop Variations

Using the authored massing and contract, the AI system first proposes and describes possible architectural directions in text. These can be discussed, rejected, adjusted, or approved before any of them is drawn.

Accepted directions are then developed into architectural propositions. In the test shown here, eight propositions (1–8) were developed for comparison. The architect or design team can select one, combine aspects of several, or reject all eight and return to the contract or generation step.

Propositions are drawn as unfinished line drawings rather than rendered images. This keeps them open to change and stops rendering quality from influencing which one is chosen.

This is the point at which computational speed has its greatest effect on the workflow. Alternatives can be explored without requiring each proposition to be manually developed to the same degree.

Architectural practice has always relied on variation, but producing alternatives has traditionally required significant time and labour. As the cost of generating variations decreases, their production becomes less scarce. The more consequential task increasingly becomes deciding which variation is worth acting on.

Generation can also produce productive deviations from the architect’s expectations. Many will simply be errors. Others may reveal relationships or formal possibilities that were not explicitly anticipated. These deviations become useful only when the architect recognizes architectural value in them; their importance lies not in the deviation itself but in the judgment that follows it.

5. Select and Intervene

Selection is an active design checkpoint rather than an approval step. The architect or design team compares the propositions and decides whether one contains a direction worth carrying forward. Selection can also revise the contract itself: a region initially left free to change may become fixed once a quality worth preserving is recognized. In the case shown here, the upper volume was initially left open in the contract but became a condition to preserve during selection.

Selection does not need to happen only through choosing an image. The architect may respond through redlines, a quick sketch over the proposition, adjustment or remodeling of the 3D geometry, combination of aspects from different propositions, or other targeted interventions. The AI system can then read the revised condition and continue from this architect-authored response. Rough sketch overlays can be useful, but their interpretation by current multimodal systems remains less reliable than clearly structured textual and geometric input.

Selection is based on what can be read directly from the propositions at this stage: composition, proportion, spatial character, consistency with the established intent, and the potential value of unexpected deviations.

A seemingly local judgment may also carry consequences that cannot be seen in the generated proposition. In Figure 6, for example, increasing a residential floor-to-floor height to 3.6 m may improve perceived spatial quality, but within the 42 m envelope established by the contract it may also reduce the total number of floors. That trade-off cannot be fully understood until the proposition returns to measurable geometry.

The annotations in Figure 6 are of two kinds. Some are new design judgments, such as fixing the upper volume. Others enforce conditions the contract had already fixed, such as the podium. If a generated proposition alters an element previously defined as fixed, that is not simply a design choice; it is evidence that the AI system failed to follow the contract and should be documented as such.

6. Reconstruct

The selected proposition is then translated back into editable three-dimensional geometry. The objective is not a blind reconstruction of the generated image, but an architectural interpretation of what should survive the transition back into a measurable 3D model.

Multimodal and agentic tools can assist in interpreting the selected proposition and translating aspects of it into modeling environments such as Rhino. Current tools do not yet guarantee dimensional fidelity or topological consistency, although their ability to support this translation is developing rapidly. The reconstructed 3D model should therefore be understood as another design state to inspect rather than as an exact or automatic translation of the generated image.

The purpose of reconstruction is not merely production efficiency. It returns the proposition to a form in which dimensions, areas, relationships, structure, environmental performance, and other architectural consequences can again be examined.

7. Evaluate and Iterate

Once the proposition has returned to editable geometry, the architect regains direct control over the project. The reconstructed 3D model can be measured, criticized, simplified, corrected, developed, or rejected. Parts of the generated proposition may survive, while others may be altered substantially once their consequences become visible in three dimensions.

This step does not necessarily complete the process. The adjusted geometry can become the starting condition for another round of generation, returning the workflow to Step 4. If the problem lies in the instructions rather than the resulting form, the architect can instead revise the contract before generating again.

In-Process AI is therefore iterative rather than linear:

Develop Variations → Select and Intervene → Reconstruct → Evaluate → Develop Variations again.

The loop continues for as long as generative variation remains useful. Once the architect determines that the project should move forward through conventional development rather than another generative cycle, it exits the loop.

After the Loop

Once further generative iteration is no longer useful, the project leaves the In-Process AI loop and continues through conventional architectural development. The geometry can be refined, coordinated, tested, and resolved before AI enters again, if useful, for materials, atmosphere, lighting, inhabitation, or other forms of visualization and communication.

This later use of AI is intentionally separated from In-Process AI. The loop ends when the selected proposition has returned to editable geometry and architectural control has been re-established; visualization becomes a subsequent representational layer rather than the endpoint of the generative design process.

What the Workflow Reveals

Where the Value Shifts

The significance of the workflow lies less in any individual tool than in how design effort and judgment are distributed through the process. Generative capacity is introduced at a point where alternatives can still reshape the architecture, while the architect remains responsible for framing the conditions, recognizing valuable deviations, selecting what continues, and deciding when the proposal must return to measurable geometry.

Once variation is no longer scarce, the value of architectural judgment shifts toward framing the problem, distinguishing useful propositions from irrelevant ones, and deciding which directions warrant further development.

What Remains Unresolved

More points of intervention do not automatically produce a better design process. The central limitation is not the quantity of alternatives that can be generated, but the information available when one of those alternatives must be selected.

As set out earlier, the architect judges propositions before their consequences can be verified. Early-stage design has always worked this way. The difference here is that some information available in the original measurable massing is temporarily exchanged for generative variation and must later be recovered through reconstruction.

Those consequences become more directly testable once a proposition has returned to measurable geometry. Structure, daylight, area, cost, and other criteria can then inform the next cycle of development. When those tests reveal a problem, the workflow loops backward so that the geometry, the contract, or the generated direction can be reconsidered.

The reconstruction step itself also remains to be fully evaluated. Current multimodal tools can assist the translation, but they do not yet ensure reliable dimensional or topological fidelity. The structural validity of the generated forms likewise still requires verification.

From Propositions to Consequences

The most important lesson from the Gaudí comparison is therefore not that AI should generate more convincing form. It is that generative systems become more architecturally consequential when the consequences of what they produce can participate in the generative loop itself.

A stronger workflow would keep more of those consequences available while alternatives are still being generated and selected. Measurable geometry would remain connected to the proposition; structural analysis could be called during iteration; daylight, area, envelope limits, cost, carbon, or other project-specific criteria could be returned alongside the generated direction rather than only after reconstruction.

The selection checkpoint would then change substantially. Instead of comparing propositions primarily through visual and spatial judgment, the architect or design team could compare each direction together with some of the consequences it produces. A decision such as raising a residential floor-to-floor height to 3.6 m could therefore be evaluated at the same moment against the 42 m envelope, total floor count, area, or other relevant constraints.

This is where the comparison with physical form-finding becomes most useful. Gaudí’s hanging models did not merely generate unexpected form; every adjustment produced a response whose consequences were embedded in the system itself. Current generative models can respond to conditions by producing propositions quickly, but architectural consequences remain partially detached from that act of generation. In the present workflow, judgment, reconstruction, and testing are still required before those effects can meaningfully inform the next iteration.

We began from the observation that AI arrives too late in the design process. The workflow reveals a second version of the same problem: consequence still arrives too late. The next development is therefore not simply better generation, but tighter integration between proposition and consequence while the architecture is still unresolved.

The Open Question

In-Process AI places generative variation inside the design process while keeping architectural judgment active at defined points. Its strongest opportunity, and its clearest limitation, both sit in the unresolved middle.

The open question is whether generated possibilities can carry enough of their consequences with them to inform architectural judgment.

References

Otto, Frei, and Bodo Rasch. 1995. Finding Form: Towards an Architecture of the Minimal. Stuttgart: Edition Axel Menges.

Pask, Gordon. 1969. “The Architectural Relevance of Cybernetics.” Architectural Design 39 (9): 494–496.

Schön, Donald A. 1983. The Reflective Practitioner: How Professionals Think in Action. New York: Basic Books.

Tomlow, Jos. 1989. Das Modell / The Model / El Modelo: Antoni Gaudís Hängemodell und seine Rekonstruktion. IL 34. Stuttgart: Institut für leichte Flächentragwerke.

The Architectural Janitor

I tend toward technological optimism, which makes this essay slightly uncomfortable to write. I have no interest in arguing for a retreat from AI, and I do not think architecture is improved by protecting old ways of working simply because they are familiar. But new technologies rarely arrive as obligations. They arrive as conveniences, then advantages, and eventually as expectations. By the time the expectation is visible, the terms of work have already begun to change.

I grew up with one hour of computer time a day. Eventually, you were expected to switch the machine off and do something else. In an All-In podcast conversation at Nvidia’s GTC in March 2026, Jensen Huang described almost the reverse condition: he would be “deeply alarmed” if an engineer earning $500,000 did not consume at least $250,000 worth of AI tokens a year. The computer has moved from something we were encouraged to limit to something a professional may increasingly be expected to keep busy. We seem to have solved the problem of screen time by making it compulsory.

Architecture has been sold this transition as a gain in time. Automate repetitive production and the architect can concentrate on design. Yet when a rendering that once took a day takes minutes, the afternoon rarely becomes free. Ten alternatives become possible, then expected. The client asks for another material, another season, their own furniture, their dog in the living room. Each request is reasonable. Refusing one can take more explanation than producing it, and the time saved by automation is gradually converted into somebody else’s expectation.

The role I worry about emerging from this arrangement is what I have come to think of as the architectural janitor. I use the term deliberately. A janitor is not the author of everything in a building, but they understand what belongs there, what has broken, what needs repairing and what should be removed. Their knowledge comes from continuous contact with the consequences of other people’s decisions. The architectural version inherits a similar responsibility: keeping a project coherent while more people and systems are able to add to it. The work can demand considerable intelligence while offering surprisingly little autonomy. You may be responsible for making everything function together while having less influence over what enters the project or what you are allowed to take out.

The client can now produce convincing images before an architect is involved. They arrive with something they like and a considerable attachment to it. Architects have long used images to build precisely this kind of enthusiasm, so we should hardly be surprised when it works without us. But when expectations have already been formed, explaining why a proposal cannot work may sound less like expertise than a failure of imagination.

Concept design is particularly vulnerable wherever it is understood as producing an attractive image. With the image in hand, the concept appears settled, even if nobody has considered what the building needs to do or whether its organization makes sense. Asking those questions can sound like charging the client to reopen a decision they have already made. It is easier to agree and continue.

Through small agreements like this, architects can lose agency without being replaced. AI accelerates an old imbalance: decisions become fixed faster than their consequences can be understood. A client’s image, a consultant’s generated report and an agent’s options can become commitments before anyone has established whether they belong together. The architect is then asked to reconcile them, with each correction requiring negotiation over something another person considers finished.

I spent years working on façades. A line that looked straightforward in an image could become a panel edge, a fixing, a drainage path and a tolerance, passing through several trades and interpretations of what had been agreed. Working through those details changed the design. Sometimes we found a better way to preserve the intention; sometimes we discovered that it depended on something nobody had resolved. Those discoveries were part of designing the building. When the image becomes an instruction, however, the same discoveries are treated as obstacles to delivering it. The architect must solve the problem while defending every departure from the picture that produced it.

This is why the reassurance that architects will retain “judgment” does not go far enough. Being asked to check a decision is not the same as being allowed to change it. The expertise needed to recognize a problem may remain valuable while the person exercising it has less influence over what happens next. A janitor who cannot throw anything out will eventually run out of space. Calling that person a curator does little to improve their position.

There is a longer consequence for the profession: the opportunities to acquire that judgment may also contract. Much of what I understand about architecture came through work whose value was not obvious while I was doing it. Drawing a junction exposed questions I did not know to ask. Revising it after a difficult conversation taught me something about both construction and the person reviewing it. Some repetition deserves to disappear, but reviewing a generated result will not necessarily provide that education. Experienced architects can draw on years of making and correcting things when they assess an output. Someone entering practice cannot simply be instructed to exercise the same judgment. If we remove those opportunities to learn, we may preserve the appearance of expertise for a generation while weakening the means of acquiring it.

My optimism rests on the possibility that these tools could let more people participate meaningfully in design and give architects more room to investigate what a building could become. A client arriving with an image can be the beginning of a productive conversation. The difficulty starts when that image closes questions it has never answered, and the architect’s appointment begins with an obligation to preserve it. That appointment needs to include permission to question the image, and time to investigate what it leaves unresolved.

The architectural janitor is a warning about accepting those terms without examining them. Being needed to make something work should give us a voice in deciding what it becomes. Otherwise, we risk becoming indispensable to resolving a project and peripheral to deciding what it is. Janitors have always held the keys. The question is whether architects will have any say over what comes through the door.

Share the Pencil

Architecture is applying AI to the wrong part of the design process. Most attention lands on what these systems can produce - images, twenty alternatives before a meeting has even started, a building rendered in seconds that once took days. But abundance of output isn't where architectural value lives, and producing more of it won't change that.

The more consequential opportunity sits earlier, in the uncertain middle of design - at the massing and schematic stages, while a building is still a hypothesis and height, volume, program, structure, circulation, and environmental response all remain negotiable. This is where variation can actually change a project rather than just describe it. Yet many emerging AI workflows do the opposite: prompt, generate, select. Ask for twenty buildings and choose the most convincing one: it looks like iteration - but the architect has quietly moved from producing the conditions of a design to auditioning finished answers, which is a different job, however similar it looks from the outside.

Removing the Architect

For the Bi-City Biennale of Urbanism and Architecture in Shenzhen and Hong Kong, working with Sandra Baggerman and Stella Zhang, we took that logic to its extreme and built a pipeline that removed human design intention from as much of the process as possible.

Hundreds of policy documents, planning texts, architectural publications and public discussions about the future of the Greater Bay Area became our source material. A custom tuned AI workflow built in ComfyUI translated those texts into architectural briefs, the briefs became images, the images became three-dimensional models, and across 15,000 iterations the system produced its own speculative landscape of possible futures: solar-optimized towers, vertical ecosystems, extreme-density settlements, synthetic wilderness, civic monuments - and, more than once, an evil-looking castle. The system could imagine any of it. It held no opinion on which future was worth building.

What struck me most was how little we actually did. We assembled what went into the system and, at the end, chose what would be fabricated for the exhibition; in between, the pipeline largely ran itself. It exposed a plausible future workflow: the architect defines the input, the machine produces the architecture, the architect returns at the end to choose. Efficient, certainly. But it is not design.

Putting the Architect Back In

It's already easy to ask an image model for something that looks like Gaudí: stone, branching columns, mosaics, curves, all appearing almost instantly. But reproducing an appearance tells you very little about the process that produced it.

Gaudí's hanging-chain models offer a better precedent. He set constraints, applied loads, and let gravity find the equilibrium. Change one weight and the whole system responds. The form was never selected from a catalogue of attractive possibilities - it emerged from an argument between the designer and a system with its own rules.

That became the premise of our more recent project, developed with Archademy, the Gaudí Foundation and Thomas Nelissen. Rather than asking AI to imitate Gaudí's architecture, we tested whether parts of his design logic could operate inside a contemporary generative workflow. The architect starts with a three-dimensional massing and fixes what matters - dimensions, program, certain geometry, a set of structural rules - and the model elaborates that massing while the design is still visibly unfinished. We intervene before material, lighting and photorealism can make one option look more convincing simply because it's prettier. The most consequential decisions in this workflow happen before the image looks finished.

A finished image invites one kind of response: I like that one. An unfinished model invites a better one: why did it do that, and what happens if I move this?

Even so, this has a hard limit. An image model doesn't understand structure the way Gaudí's hanging chains understood gravity: gravity has consequences, a generative model has probabilities. A generated form can resemble one produced by structural logic without possessing that logic, and that gap is exactly where architectural judgment, and responsibility, still has to live.

Staying in the Middle

Prompting won't be the defining AI skill for architects - prompts will keep getting easier, models more capable, interfaces less visible. The harder skill is knowing where to intervene: site boundaries, program, structural grid and maximum height fixed before the machine ever runs, the façade and internal subdivision left open for it to explore, and the discipline to resist its proposal when it drifts past what should still be negotiable.

That's what sharing the pencil actually means: staying inside the process long enough that each move changes what happens next, rather than handing over a brief and waiting for a result.

As architectural output gets cheaper, authorship will increasingly belong to whoever sets the conditions those outputs emerge from, and that reaches beyond how architects work at their desks. In New York, as elsewhere, repeated building types make up most of the city we actually inhabit, and the conditions set upstream determine what gets repeated downstream.

Adapting to AI means deciding what the machine isn't allowed to touch. The unresolved middle remains architecture's most consequential territory - for as long as there's something left to design.

The Recursive Data That Shapes Our Built Environment

The integration of artificial intelligence into professional practice is poised to exert a transformative influence on the built environment, extending beyond mere automation to reshape the frameworks within which we operate. Legal codes may soon be drafted with assistance from large language models, urban strategies distilled through machine-generated analysis for urban planners, and architectural concepts informed by AI-generated imagery that will eventually manifest as tangible structures. As we see, artificial intelligence does not merely produce—it interprets and refines its own outputs, reincorporating them into subsequent iterations. The designs we generate today are thus products of datasets whose origins are largely human made, but these outputs will, in turn, inform the training of future models. Creating the start of the recursive dataset where AI start to inform itself. What are the implications when artificial intelligence both authors and evaluates the narratives that define our urban landscapes, constructing the future through a self-referential exchange largely opaque to human oversight?

The datasets underpinning these models appear vast and generative, offering a breadth of innovative forms and conceptual possibilities. However, this apparent abundance obscures a critical limitation: their composition is neither fully transparent nor necessarily exhaustive. Have certain architectural typologies, material traditions, or regional practices been excluded—whether by design or oversight—prior to their availability for practitioner use? Claude Shannon, a mathematician at Bell Labs instrumental in establishing modern computing, envisioned information as a universal system capable of encoding all knowledge with precision. In contrast, Timnit Gebru, a computer scientist renowned for her contributions to AI ethics, demonstrates that datasets are not neutral repositories; they are curated constructs, shaped by decisions that often evade scrutiny. The creative potential of artificial intelligence is thus contingent upon the unseen boundaries of its training data.

This tension is not without precedent. Michel Foucault, a French philosopher and historian, argued that power is exercised not only through what is preserved but also through what is excluded from the record. In the mid-20th century, institutions such as the Bouwcentrum Rotterdam exemplified a commitment to accessible knowledge, providing architects and planners with open access to data on materials, zoning regulations, and urban planning methodologies. Such transparency enabled practitioners to discern both presence and absence within the corpus. Today, this openness has largely eroded, supplanted by proprietary datasets embedded within artificial intelligence models that resist examination. Not just the big tech companies are doing this—most firms sit on large troves of data, saved on servers somewhere, completely unsearchable and hoarded away from the communal greater good. James Bridle, a British writer and artist, terms this phenomenon the "new dark age"—an era where technological complexity proliferates, rendering the mechanisms of control increasingly inscrutable.

Consider the iterative nature of contemporary design processes. An artificial intelligence model, trained on existing architectural precedents—perhaps a sleek, parametric structure—generates an image that informs a subsequent project. This project is archived, and a future model incorporates it into its training, perpetuating the pattern. Should the initial dataset omit vernacular architectures, informal settlements, or non-Western traditions, these elements may gradually vanish from the disciplinary lexicon. Unlike historical canons curated by scholars or critics, this emergent canon is sculpted by the black box hierarchies of machine learning, subtly redefining the scope of design potential without explicit deliberation.

This selective memory extends to the preservation of architectural heritage. Envision a historic district under consideration for conservation. An artificial intelligence system analyzes photographic records and archival data, prioritizing features deemed “significant”—perhaps ornate facades—while overlooking less conspicuous elements, such as utilitarian outbuildings, due to biases embedded in its training. A subsequent model drafts a report for municipal review, further eliding these overlooked aspects. By the time a decision is reached, the outbuildings are excluded—not through deliberate rejection, but because the artificial intelligence’s concealed preferences shaped the narrative. Foucault’s insight into the power of exclusion finds a modern echo here: what is remembered is determined not by chance, but by the parameters of recognition.

The phenomenon intensifies as artificial intelligence engages in self-dialogue. A recent anecdote illustrates this: a student provides ChatGPT with bullet points to compose an email, and a professor uses another instance of ChatGPT to reduce that email back to bullet points—an exchange entirely mediated by machines. More consequential is the case of DeepSeek, an artificial intelligence model reportedly trained not on primary textual sources but on outputs pre-processed by ChatGPT. Rather than accessing unfiltered documents, DeepSeek learns from an abstracted layer of machine interpretation, akin to a secondary distillation. In architectural practice, this suggests a scenario wherein one model generates a design, another summarizes or modifies it, and the process continues indefinitely. Benjamin Bratton, a philosopher and professor at the University of California, San Diego, posits that such computational systems constitute an emergent form of governance—one arising not from explicit legislation but from the accretion of algorithmic interactions.

This interplay is palpable in practical applications. An architect employs artificial intelligence to draft a city block; a second model interprets the proposal, highlighting features such as high-rise structures or public plazas, and produces a summary; a planner’s model then adjusts the plan—perhaps increasing parking at the expense of greenery—based on that interpretation. Each stage involves artificial intelligence communicating with its counterparts, with human input relegated to a supervisory role. The initial vision of a verdant, pedestrian-friendly space may devolve into a utilitarian layout, altered not by intent but by successive machine-mediated reinterpretations.

Such dynamics are not confined to large-scale projects. Consider a municipal effort to redesign a public park. An artificial intelligence system compiles data—visitor counts, climate metrics—and proposes a layout with benches and pathways. A subsequent model condenses this into a report for civic authorities, omitting shade trees not deemed “noteworthy” within its dataset. A contractor’s model then revises the plan for cost efficiency, eliminating a playground. The final park—a sparse expanse with minimal features—bears little resemblance to the original intent, reshaped through a chain of artificial intelligence exchanges unnoticed by its human overseers.

The implications resonate in sustainable design as well. An artificial intelligence tool tasked with optimizing a building’s energy performance might analyze insulation data and climate patterns, recommending a glass-heavy façade based on prevalent trends in its training set. Another model, assessing lifecycle costs, adjusts the proposal—favoring cheaper materials over long-term efficiency—because its data prioritizes initial expenditure. The result may undermine sustainability goals, not due to flawed intent, but because the artificial intelligence’s hidden preferences skewed the outcome. Practitioners, unaware of these biases, risk endorsing solutions misaligned with broader ecological imperatives.

Nevertheless, the potential of artificial intelligence remains striking. Its datasets are not inert; they pulse with possibilities, yielding designs that defy conventional imagination. The challenge lies not in their capacity, but in our inability to ascertain their scope. An exploration of an AI-generated urban dataset revealed a diverse array of spatial configurations—yet it offered no indication of excluded elements. Research into search engine algorithms reveals a parallel: even in contexts where information access is presumed robust, content is filtered by opaque criteria. Practitioners accustomed to transparent frameworks may thus remain oblivious to the boundaries circumscribing their tools.

Peter Weibel, an artist and professor at the University of Applied Arts Vienna until 2017 (and later director of the ZKM Center for Art and Media Karlsruhe), proposed a “Datatopia”—a vision wherein data is not hoarded but collectively interrogated and reconfigured. For architects and designers, this suggests a reimagining of agency: not merely to deploy artificial intelligence, but to probe its foundations. What narratives are embedded within the datasets that define our built environment? What possibilities remain unarticulated, and who—or what—determines their salience? As artificial intelligence increasingly mediates both the conceptualization and appraisal of design, these inquiries extend beyond technical utility to the epistemology of knowledge in a machine-driven era. Should we advocate for a concerted effort to unveil these systems, demanding transparency as a prerequisite for their integration? Or might we adopt a stance of cautious observation, allowing this interplay to evolve while we assess its trajectory? The question persists: as artificial intelligence converses with itself to construct our world, are we prepared to assert our role within that dialogue, or will we acquiesce to a future narrated by mechanisms we neither fully see nor comprehend?