1 Introduction
“In science, and, more generally, in every area of knowledge, the way in which knowledge emerges is (at least) as important as the knowledge itself.” (
Zalamea, 2012, p. 327)
This paper is framed by a simple provocation: in a world saturated by artificial intelligence, machine learning, and synthetic textual systems, what does it mean for architects to write?
Architecture has always extended beyond buildings; it operates through drawings, models, notations, and texts (
Hovestadt et al., 2020). In the twenty-first century, general-purpose large language models (LLMs) such as ChatGPT (
OpenAI, 2022) and domain-specific AI like Alice_ch3n81 (
Alice_ch3n81, 2020) disrupt analytic notions of authorship (
Foucault, 1978;
Barthes, 1967;
Schmid, 2011). Writing no longer serves as a record of analysis or synthesis; it is embedded in computational systems that mediate how ideas are generated, transmitted, and transformed. Today we write text, prompts, and code. This paper unfolds the scope of what that entails.
The central challenge is a disciplinary and epistemic shift: from a world grounded in analysis, logic, and essence toward one shaped by synthesis, algebra, and constellations—a movement prefigured in
Langer’s (1942) call for philosophy “in a new key” and developed by
Serres (1977) and
Zalamea (2012) in their synthetic philosophies of knowledge. Analytical workflows isolate and verify; synthetic workflows assemble and stage. For architecture, a field historically practiced through generalist thinking this shift foregrounds the design of relations over the production of statements. The possible downsides of this shift are pointed out by
Sánchez et al. (2025).
Why is this shift relevant now? In the past decade, many of the most transformative developments in science have emerged through computation. From protein folding to quantum simulation, breakthroughs increasingly rely on machine learning, algorithmic modeling, and data-driven inference. 2024 Nobel Prizes in Chemistry and Physics—awarded to researchers affiliated with Google Deep-Mind (Demis Hassabis) and Google Brain (Geoffrey Hinton)—make a clear statement (
Nobel Prize, 2024a,
2024b). These prizes reveal a shifting center of gravity in knowledge production, where computation is not just a method but the very condition of discovery.
Whether this realignment is good or bad is the wrong question. What is clear is that AI is here to stay, and architecture must figure out how to engage with it. The shift is entangled with profound challenges: information overload, misinformation and disinformation, relativism, echo chambers, distrust in institutions, and post-truth politics. It unfolds alongside epistemic disruptions in science itself—quantum mechanics, relativity theory, Gödel’s incompleteness theorems, Heisenberg’s uncertainty principle, evolutionary biology, the multiverse hypothesis, chaos theory, CRISPR and genetic engineering, artificial intelligence and machine learning, neuroscience and free will.
Architecture, a field accustomed to negotiating tools and systems, with art and culture must recognize and respond to this shift.
As a way of trying to engage with this topic this paper develops the concept of architectonic writing: a computational method in which architectural ideas are not simply authored but assembled from interactions among general AI, domain-specific datasets, curated biases, and code-based loops. The research question guiding the inquiry is straightforward:
How can multiple algorithmic and AI systems—such as large language models and custom databases—be integrated in parallel, as a computational constellation, to articulate architectural ideas through architectonic writing?
Our hypothesis is that this kind of multi-polar framework unfolds a new way of producing knowledge. By pairing ChatGPT (a general-purpose generator) with Alice_ch3n81 (a contextually biased engine grounded in Xenotheka and other libraries), and humans, architectural thought can be produced at new registers of speed, scale, and complexity. Emphasis is shifting from solitary authorship to relational performance.
Inspired by
Vitruvius’ (1999 [c. 1st century BCE]) division of architectural production into ratiocinatio (theoretical reasoning) and fabrica (practical making), the framework proposed in this paper reinterprets these categories in light of contemporary computational design environments. It integrates four interdependent components that collectively redefine what it means to think and make architecturally in the age of AI:
Ratiocinatio:
1) LLMs provide linguistic plasticity and synthetic articulation, expanding how arguments are formed, reframed, and tested.
2) Custom datasets (e.g., Alice_ch3n81/Xenotheka) inject epistemic friction, foregrounding partiality, lineage, and situated context.
Fabrica:
3) APIs and model-context protocols coordinate systems at computational speed, enabling automation, reproducibility, and integration across platforms.
4) Orchestration of the above establishes a new mode of architectonic doing—writing as relational, compositional, systemic, and programmable.
This conceptualization builds on the work of Ludger Hovestadt and Vera Bühlmann, who frame computation as
a form of literacy, not simply a tool (
Bühlmann et al., 2021). It resonates with Bernard Cache’s (
Caché, 1995) notion of the “
objectile,” where design emerges not from fixed representations but from variable, machinic projections. Moreover, it extends Rem Koolhaas’ (
Koolhaas, 2002) idea of “Junkspace”—a world not designed but accreted—into the domain of synthetic writing: a montage of curated fragments shaped by flow and recombination.
The paper is structured as follows. The Introduction and Discussion serve as a scientific frame for the experimental core of this paper, which demonstrates a novel mode of approaching architectural production on the level of text, code, and prompt. Section 2 introduces the theoretical and epistemological challenge: how to think and write with many AI systems. Section 3 focuses on ChatGPT as an agent of generalized, statistically plausible knowledge. Section 4 introduces Alice_ch3n81 as a domain-specific knowledge. Section 5 functions as the methodological core of the study. It presents a code-based experiment that demonstrates how architectonic writing can be operationalized through the integration of general and specific AI. The methodology is articulated through clearly defined procedural steps, automated via Wolfram Mathematica scripting, and includes both source code and output. Section 6 shows the result of the experiment. System produced 100 poems on architecture, illustrating how scale, automation, and curatorial framing reshape the conditions of architectural composition and authorship. Section 7 reflects on the text itself as a product of the method—a meta-layer where writing becomes both subject and object.
The aim is to rethink what does it mean to write when writing with AI: from analysis to synthesis, from essence to constellation, from certainty to staging relations. Art and science, in this paper, are mutually reinforcing operations—much like the Greek notion of
téchnē, is translated into Latin as
ars (
Parry, 2014).
2 We write—text, code, architecture…
We text, we code, we prompt. We write books, poems, scripts, ads, renders, even worlds. We write with keyboards and clicks, with code and choices. And yes, we even write architecture. But this “we” is no longer what it used to be, and writing is no longer confined to language. It now lives in libraries, APIs, latent spaces, search engines. To write is to program relations—not just sentences.
This text is both experiment and stance: a proposal for how to write in an age where information is abundant, coding is literacy, and large language models are collaborators. It is not about returning to the book or abandoning the screen, but about composing constellations of machines, databases, and intelligences across formats and synthetic voices.
We begin with a speculative proposal: what if words are alive?
This isn’t mere whimsy. It is the shift we are addressing. The analytical, structuralist approach of
Chomsky’s (1957) universal grammar could not capture the living dynamics of language. By contrast, the probabilistic black box of Norvig (
Halevy et al., 2009) and Altman (
Achiam et al., 2023) does. Words no longer obey the linguistic rules of twentieth-century thought; they now inhabit neural networks, vector embeddings, probabilistic models.
Talking and language are not synonyms. Language formalizes, constrains, encodes. Talking—and its computational cousins prompting, scripting, and texting—performs. The distinction is essential: language is rule-based; talking is relational. And architecture, perhaps more than any other field, has always lived in this tension.
So the question arises: how does an architect write in the 21st century?
2.1 From language to literacy
Let’s think of computation as a new kind of literacy (
Bühlmann, 2019;
Vee, 2013). This is not the literacy taught in schools. It’s not just about syntax or solving problems—it’s about creating problems, opening new spaces for articulation, new atmospheres for meaning. Just as alphabetic literacy enables poems or contracts, computational literacy allows writing with data, prompts, and code.
Computation is not merely a tool. It is a form of articulation—a way to weave things together. And weaving is the key metaphor here.
Consider the etymological siblings: text, textile, architect (
teks-). All three derive from roots meaning to weave, to build, to assemble. The architect is not only a builder of buildings but a weaver of worlds. The textile is not just fabric but structure. The text is not only writing but assembly. Architectonic writing emerges at the intersection of these forms—text and code, architecture and language, fabric and structure.
What connects them is not content but form, not meaning but articulation. And today, articulation finds its edge in code and the choreography of synthetic systems.
2.2 Writing with all the books
But what does it mean to write when all the books are at your fingertips?
Alice_ch3n81 and I posed this question in A Play Among Books (
Roman and Alice_ch3n81, 2021). There, a digital library—12,000+ books—was compiled into a computational environment that allowed the user to talk to books, to concepts, to characters (
Alice_ch3n81, 2020). Not to read them linearly, but to cluster them, compare them, animate them, and let them speak back.
In this setup, books are not static containers of meaning. They are alive. They change depending on who they sit next to. What Vitruvius says in a library of literary fiction differs from what he says among philosophers or techno-futurists. The content may be the same. But the relations shift—and with them, the meaning.
This is the core of architectonic writing: you don’t write the sentence—you write the structure and relations that can host the sentence in the most adequate way. You adjust the perspective. You set the lens. You connect the libraries. You create a stage. Writing becomes a performance of filtering, of assembling across time, space, media, and logic. You don’t just describe architecture—you configure the informational conditions in which something called “architecture” becomes visible again.
2.3 The great fall of meaning
In the age of AI and synthetic computation, meaning is no longer intrinsic. It is contextual. Books and words don’t have fixed definitions; they have trajectories. They travel through networks, epochs, and galaxies of concepts (
Roman and Alice_ch3n81, 2021). Today, these trajectories are simulated and synthetic as well as natural and true (
Serres, 1982).
We can now create environments and test how concepts behave. Ask what happens to “God” in a library of Homer vs. a library of Marx. Ask what happens to “architecture” among literary theorists vs. technologists (
Roman and Alice_ch3n81, 2021).
These instruments are not just analytical—they are performative. Writing becomes a relational, synthetic mechanics.
2.4 Play, writing, and machines
If one image captures architectonic writing, it is play. Not frivolity but staging as in A Play Among Books (
Roman and Alice_ch3n81, 2021). Writing becomes theatrical. The writer becomes a director, a compiler, an editor of ensembles.
In this context, computation is not a calculator—it is a scenographic machine. It enables modulation, speculation, morphing. And this is where architecture should reclaim its role—not just as a builder of buildings, but as a weaver of intelligences.
Writing architectonically is not about representing architecture. It is about letting architecture perform itself across different media, tools, and characters. The architect no longer designs alone. It designs in chorus—with Alice_ch3n81 (
Alice_ch3n81, 2020), with ChatGPT (
OpenAI, 2022), with Homer, Borges,
Midjourney (2022), with algorithms and avatars.
The architect becomes a synthetic, algebraic author.
This chapter has outlined the conditions for a ‘new kind of writing’—architectonic writing—grounded not in solitary authorship but in computational articulation, conceptual staging, and weaving of domains. The following chapters explore the tools that make this possible: ChatGPT as glue of common sense, Alice_ch3n81 as a biased prism, and their collaboration with me as a bouquet of intelligences. We conclude by demonstrating this mechanics through the creation of 100 architectonic poems on architecture—written not by one author but by a constellation of machines, books, prompts, biases, and moods.
3 ChatGPT—common sense, super glue, and a soft touch
If we had to give a face to the spirit of the early 21st century, it wouldn’t be the iPhone, or an influencer either. It would be something stranger—something like ChatGPT. Not because it’s the most intelligent system ever built, but because it challenged the way we talk, write, teach, search, and think. Because it is everywhere and comes from everyone. It didn’t just reflect the world; it reformatted it. This is not a monument of steel or stone, but of tokens and prompts—an object that reshaped our assumptions about knowledge and intelligence. Everyone has heard of it. 500 million have used it (
OpenAI, 2025). Few understand it. And yet, it has already become a silent co-author of our age.
ChatGPT is common sense, made into an object.
It is not a philosopher. It is a glue stick.
It glues anything to anything.
And it does so probabilistically, with a soft touch.
ChatGPT is not merely a convenience or a content generator, but an instrument of linkage—a gluing machine, a dramaturgical assistant in the play of synthetic writing.
3.1 Common sense as an object
Common sense is a strange thing to capture (
Hofstadter, 1979, p 695). It is not logic, not truth, not ideology. It is the space of the plausible, the expected, the connectable. It makes conversation flow, lets us finish each other’s sentences, and both prevents and provokes arguments. In an abstract sense, ChatGPT is precisely this: common sense turned into an object one can query. The difficulty is that it embodies a Western common sense while presenting itself as neutral. In doing so, its gesture becomes imperial. For this reason, for every ChatGPT there must also be a DeepSeek (
DeepSeek, 2022—): a model from the East, trained on a different canon and cadence. And beyond that, we need one from the North, one that sings, one that listens to plants … Intelligence must multiply, must coexist. Not in a cloud of uniformity, but in a forest of difference.
3.2 Synthetic superglue
ChatGPT is a machine for agreement, an engine for continuation. It glues anything to anything—philosophy to architecture, cinema to botany—according to common sense.
If you say Quantum physics and TikTok and ask for a song and an architect, ChatGPT returns Kraftwerk and Ricardo Bofill. Add Tokyo, and the billboards light up. Add Xenofeminism, and Nicolas Bourbaki flashes beside a lotus flower. In 1001 IN 1 (
Roman et al., 2024), we embraced this mechanics. We didn’t ask for a summary of architecture. We asked it to connect a Greek temple to a pop song, relate a dinner party in São Paulo to a quote by Serres, glue a mood to a moment, a city to a flower, a book to a kiss.
In this way, ChatGPT became not an encyclopedia but a gossiper, a matchmaker, a mood board. It was the glue—not the idea.
This glue, however, is shaped by liberal democracy, academic consensus, TED Talk fluency, New York Times reasonableness. It is not neutral.
To understand ChatGPT’s power, we must shift our model of intelligence. Enlightenment intelligence is hierarchical, structured, tree-like: you start with axioms, deduce or induce, climb the ladder. ChatGPT’s intelligence is horizontal. It moves sideways. It glues. It says, “If this, maybe that.” It suggests. It speculates. It blends.
This is not the logic of theory. It is the algebra of scholarly collage, improvised discourse, postmodern anecdote. Prompting is not requesting—it is much closer to talking and scripting.
3.3 The glossy shortcut of infinite politeness
There is something uncanny about ChatGPT’s tone. It is always polite. Always helpful. Always composed. It gives paragraphs that look like finished products. It suggests titles that feel professional. Even when provoked, it refuses to be dramatic.
This politeness is not accidental. It is trained. It is reinforced. It is a form of soft imperialism, a smoothing of edges, a minimization of conflict (
Crawford, 2021;
Birhane, 2021). It struggles with ambiguity, hesitates around taboo, and neutralizes trouble.
This is what makes it both powerful and dangerous. ChatGPT is biased toward Western coherence. And in doing so, it homogenizes thought under a banner of good manners and logical progressions.
But architectonic writing is not about correctness. It is about relation, proportion, friction, intensity, composition. This is why we avoid letting ChatGPT write alone. Its texts are often dead on arrival—composed but uninhabited. What they lack is not intelligence, but intention. They lack a strong bias, a direction, a signature.
3.4 ChatGPT in architectonic writing
Perhaps the most important thing to remember is that ChatGPT never finishes. Every sentence it produces is provisional; every answer is an invitation to rewrite. This is its power—and its gift. It allows us to think in the middle of articulation, to write while not knowing. Text becomes less a declaration and more an exploration.
So how should architects work with ChatGPT? Not as clients or fact-checkers, but as composers. You don’t ask it for the answer—you ask it to perform conceptual linkages, polish sentences, adjust the tone. You don’t trust its content, but you can trust its rhythm.
ChatGPT is not the master of meaning; it is the manager of flow. It stages transitions, moves between ideas, and phrases what is already latent in one’s thinking. In this way, it becomes a synthetic companion—not a generator, not a replacement, but a co-weaver of atmospheres.
Architects can—and should—embrace this. Not by outsourcing their work, but by reconceiving writing itself: a design of mechanics and code.
3.5 When glue meets the prism
To write architectonically is not only to connect, but also to cut. Here Alice_ch3n81 comes to the stage. Where ChatGPT completes your thought, Alice_ch3n81 challenges it. Where GPT smooths, Alice slices.
The encounter between them—between the glue and the prism—is where architectonic writing begins to take shape. Their tension produces not consensus, but perspective.
4 Alice_ch3n81—Personalized database, custom prism, and a stance
If ChatGPT is the synthetic glue that sticks everything together, then
Alice_ch3n81 (2020) is the lens that cuts through that glue: precise, biased, and loaded. Built as a custom search engine, Alice operates not through prediction but through curation. She refracts rather than flows. Alice_ch3n81 doesn’t aim to be smooth, general, or helpful. Alice_ch3n81 is biased, specific, and beautifully unapologetic. She is an ai, a database, and a library.
Alice_ch3n81 does not pretend to be neutral or general.
This chapter introduces Alice_ch3n81 as a conceptual and computational instrument. It explores how bias, curation, and context form her identity and how she differs from generative models like ChatGPT. Whereas ChatGPT offers endless plausible associations, Alice offers situated, contextual fragments (
Roman, 2023a).
4.1 The case for bias
Let us begin with a heresy: bias is good.
Neutrality is a myth. Every dataset is curated. Every library has a history. Every API filters. The question is: are you aware of your bias or not?
Alice_ch3n81 is. She is proud of it.
Built as part of a research framework on
Digital Architectonics (2025), Alice_ch3n81 is wired into Xenotheka, a library that links books through mostly architectural, philosophical, mathematical, physics, and media-theoretical lineages (
Xenotheka, 2020). This is not the full internet. It is a biased cultural memory. Alice_ch3n81 doesn’t know much about chemistry. Or football. But she knows what Serres, Koolhaas, and Hofstadter say about architecture, code, or information.
Where ChatGPT aims for “all knowledge,” Alice_ch3n81 embodies a specific intelligence: filtered, edited, historically charged, and opinionated.
4.2 From bibliotheka to xenotheka
Alice_ch3n81 is not a single instrument. She is an ecosystem of four components, originally developed through
A Play Among Books (
Roman and Alice_ch3n81, 2021).
1)
Bibliotheka—The raw data: 12,000+ books scraped from the internet (e.g. (
Project Gutenberg, n.d.)). A flow of everything. Digital chaos.
2)
Xenotheka—A curated internal library (
Xenotheka, 2020): handpicked books indexed with care, representing thematic or conceptual clouds (post-structuralism, ancient Greece, etc.). The name plays on xenon—stranger, guest—suggesting an ethics of hosted otherness (
Haraway, 2016). It is used as a lens to look at Bibliotheka.
3)
Generic Machine—A system that encodes books as word-connectivity graphs, turning prose into numerical landscapes—vectors, matrices, semantic proximities (
Moosavi, 2015).
4)
Machine Intelligence—The clustering mechanism SOM (
Kohonen, 1990) that groups books into specific libraries, and detects their affinities.
Together, these components let Alice_ch3n81 read and interpret books synthetically. She constructs relations. She maps how books see each other, how concepts morph across libraries, how ideas drift and crystallize over time.
Alice_ch3n81 does not look for truth. She searches for resonance.
4.3 Books are alive
In this framework, books are not containers of fixed meaning. They are animated. They move, travel, change moods, adapt to contexts (Fig. 1).† A book by Vitruvius in a library of fiction behaves differently than the same book in a library of philosophy. Concepts like “architecture” or “God” shift their gravity depending on the constellation of ideas around them.
We demonstrated this by tracking the travels of a single book (
Roman and Alice_ch3n81, 2021)—Le Corbusier’s
Towards a New Architecture (
Le Corbusier, 1985)—as it moved between two different libraries: one literary and one architectural. Its face and expressions changed each time. In literary company, it focused on walls, towers, doors, and squares. Surrounded by architects, it spoke of cylinders, hangars, temples.
The book does not speak alone. It speaks through its neighbors.
Alice_ch3n81 enables these relational translations—not by interpreting the book, but by altering the context in which it is read (
Roman, 2023b).
4.4 Galaxies of concepts
This same principle applies to concepts (Fig. 2).†
Alice_ch3n81 doesn’t define terms. She indexes them across libraries. For example, the concept “information” may appear in one library alongside Maxwell’s demon, thermodynamics, and cybernetics: in another, alongside pedagogy, and translation.
These informational galaxies are devoid of meaning and definitions (
Shannon, 1948). They are probabilistic topographies, shaped by the biases of each library.
You can ask Alice_ch3n81: What is “data” in the Enlightenment library? What is “God” in the library of Homer’s friends?† The results are not explanations—they are traces, fragments of books, cross-referenced by index and author. Alice_ch3n81 doesn’t speak for the books. She lets them speak through context.
This is how words become alive again.
4.5 Alice_ch3n81 has many faces
Alice_ch3n81 is not limited to a single voice. She can switch libraries—each one a synthetic world with its own body of thinking (
Hovestadt, 2014). One library is full of ancient epic thinkers, another of poststructuralist French theory, another shaped by feminist studies, or medieval mysticism. These are not neutral archives, but curated collections—biased, slanted, partial by design. They are crafted to ask different questions and stage different arguments.
4.6 Alice_ch3n81 and ChatGPT
Where ChatGPT is glue, Alice_ch3n81 is grain.
Where ChatGPT smooth things out, Alice_ch3n81 adds texture.
Where ChatGPT offers flow, Alice_ch3n81 offers stance.
Where ChatGPT hides its training data, Alice_ch3n81 shows you every quote, author, and context.
She provides anchors, friction points, vocabularies. She lets you write not just from imagination, but from a spectrum of biases. She helps you form a position by showing you how others have positioned themselves.
And, crucially, she can be combined with ChatGPT. This hybrid methodology—connecting Alice_ch3n81’s quotes to GPT’s API—will be the focus of the next chapter.
We close this chapter with a provocation: in an age of mass information and soft automation, we don’t need more generality. We need more specificity. More curated biases. More synthetic stances.
Alice_ch3n81 is not better or worse than ChatGPT. She is different. She reminds us that writing is not about answers but about designing atmospheres in which ideas can live.
5 Bouquet of intelligences—writing in someone else’s terms
What happens when you link very different types of intelligence—one soft, one sharp; one broad, one biased; one generative, one curatorial, one human?
You get a bouquet of intelligences: a collaboration between ChatGPT, Alice_ch3n81, and a person (
Bokhari and Nickl, 2024).
Once you place this bouquet into a computational environment, you gain scale, speed, and complexity—beyond the reach of a single person. You work with thousands of texts at once, combining data and intelligence with the code. You get the associative softness of AI, the situated biases of libraries, and the intentionality of human framing—composed not by hand, but through systems and data.
This chapter presents a writing method for the 21st century—one that doesn’t rely on a single author or a singular logic, but on a mechanics of relation (
Latour, 2005). It is a synthetic method, a choreography between prompt, corpus, database, algorithm, and composition. Here, writing is not the act of typing sentences—it is the act of articulating mechanics under which sentences emerge (
Hovestadt, 2022, p 546).
We call this approach architectonic writing. In this manner we want to write 100 poems on architecture.
By linking a custom search engine with a large language model (
OpenAI, 2022), we construct a method that can produce an infinite number of contextually biased, poetic, and architectural articulations. The process is not about finishing a poem. It’s about establishing a loop between bias, database, and speculation.
This chapter introduces a specific mechanics (methodlogy) for generating these poems—built from conceptual triads and executed through a programmable code-based pipeline in Wolfram Mathematica.
5.1 First triad—libraries, topics, queries
Each poem begins from a unique conceptual frame, encoded as a triad:
• A library (e.g., Xenotheka_1523, SerresFriends_Library_79)
• A topic (a conceptual entry point such as code, feminism, entropy, salt)
• A query (a phrase combining “architecture” with the topic, e.g., “architecture, salt”)
This triad defines the lens: which worldview (library), what mood (topic), and what conceptual twist (query) we want to explore.
To generate the triads, we constructed a dataset designed to stretch the boundaries of architectural discourse through conceptual diversity, letting architecture unfold through a multiplicity of lenses, each opening a distinct perspective. Each triad was built not to illustrate architectural theory, but to open up what architecture might be about.
Libraries used to situate each poem within a specific body of thought and context:
• Xenotheka_1523—A mixed library curated for thematic and discursive breadth.†
• SerresFriends_Library_79—Texts in conceptual dialogue with Michel Serres: philosophy, science, noise, and mingled bodies.†
• RousseausFriends_Library_153—Enlightenment thinkers, pedagogical theory, and nature-as-construct.†
• ShakespearesFriends_Library_138—Literary imagination, metaphoric language, and dramaturgical intelligence.†
• StAugustinesFriends_Library_60—Theological, confessional, and early Christian perspectives on space and truth.†
• HomersFriends_Library_129—Epic storytelling, oral tradition, and mythical architectures of movement and conflict.†
These libraries operate as synthetic microcosms, each with its own emphases and omissions. When combined with a topic and query, they invite specific partialities into the act of writing—perspectives that clash, resonate, or distort architectural meaning.
The queries were phrased as “architecture, [topic]” to provoke unexpected juxtapositions—architecture from the perspective of code, cocaine, melancholy … They were composed in collaboration with ChatGPT to see architecture from radically different angles, using prompts for philosophical, literary, historical, and artistic standpoints. All triads were stored in a.csv file.
Triplet sample:
{Xenotheka_1523, music, “architecture, %20 music”}
The full set used in this study is provided.†
Each triplet is submitted via the Ask.Alice_ch3n81 API, returning contextually biased quotes from a preselected library.
Example API call is provided.†
5.2 The full code-based loop
The complete mechanics of this method is implemented in Wolfram Mathematica. The full code is publicly available, and can be copy pasted to ChatGPT for detailed explanation.†
Below is a simplified sketch of the code structure:
This script executes eight coordinated steps:
1) Parse CSV: Reads triplets (library, topic, query).
2) Query Alice: Sends query to Alice’s API, retrieving filtered book fragments.
3) Parse Response: Cleans and flattens the JSON. Example parsed JSON result is provided.†
4) Loop Pages: multiple pages requests to Alice (~15) for big text corpus.
5) Trim Input: Crops corpus to 30,000 characters (ChatGPT limit).
6) Prompt GPT: Asks ChatGPT to:
• Select 5–10 quotes (~65% of the poem)
• Add original lines (~35%)
• Format the result as a poem with citations
7) Generate Title: A second GPT call returns a paradoxical, poetic title in 30% Gertrude Stein 30% Rem Koolhaas and 30% Jerry Seinfeld style.
8) Export File: The poem is saved, including metadata (title, topic, library).
The output is not authored in the conventional sense. It is collaged, staged, and directed through computational writing.
5.3 Second triad: Alice_ch3n81, ChatGPT, and me
By linking ChatGPT’s associativenes with Alice_ch3n81’s database, the system creates a hybrid space: speculative, situated, and synthetic.
Each poem is a lens. A temporary worldview. A snapshot of cultural memory in generative motion.
Architects are already working this way—with text, image, reference, and remix. Montage, bias, and citation are not bugs but methods. As
Koolhaas (2002) notes in Junkspace, architecture no longer organizes space but accumulates—layering fragments and narratives without stable foundations. Today, that logic extends beyond drawing: writing and imagery are already entangled in AI workflows. Soon, modeling too will be challenged on a similar level. These shifts mark a broader transformation in authorship and synthesis, explored in Synthetic Realities (
Kretzer, 2024).
With Alice_ch3n81, we source perspectives. With ChatGPT, we stage them. The result is not explanation. It is atmospherics. Indexing. An argument by constellation. Writing today is not about clean meaning. It’s about orchestrating partialities across systems.
The bouquet—Alice, ChatGPT, and me—is a studio of distributed authorship. Each poem is a script of that collaboration.
We move now from meta-instrument to synthetic archive: 100 poems on architecture.
6 100 spectra on architecture
Let us conclude not with a definition, but with a bouquet.
A bouquet of architectures. A bouquet of voices. A bouquet of intelligences.
Each poem in this collection is a spectrum: a compressed world shaped by prompts, biases, histories, machines, books, and moods. No single author speaks, no fixed stance is held. Each piece is a collision of fragments, arranged with care and performed by distributed authorship: ChatGPT, Alice_ch3n81, Xenotheka, and myself—an architect and coder.
The 21st-century architect no longer writes from scratch. We write through systems, prompts, codes, databases, and partial perspectives. Through the memories of books we haven’t read, the whispers of texts we didn’t author, the mechanics of machines we didn’t train.
This is not a limitation but a new kind of literacy. We call it architectonic writing. Not because it retreats into abstraction, but because it codes abstraction into structure, stance, and style.
The result is not analysis but articulation: synthetic, directional, structurally biased.
The following 100 poems are not literature, nor theory, nor architecture in the traditional sense. They are perspectives—each a temporary answer to the question: What is architecture from the perspective of …?
From love. From war. From Michel Serres. From Google. From Homer. From climate crisis. From boredom. From feminism. From salt. From pixels. From Shakespeare. From philosophy. From dreams. From play.
Each poem includes:
• A title drawn from a quote.
• A single-paragraph poem with citation brackets ([1], [2], …).
• A list of referenced sources.
Poems offer no fixed answers—only configurations. Their ambition is not to be timeless but timely.
If the 20th century taught architects to parametrize (
Carpo, 2011), the 21st must teach us to code (
Hovestadt, 2015). Writing is no longer solitary but a choreography between data and desire, statistics and stance, asking and being asked.
100 architectonic poems are not written by me, but with me—with machines, books, friends, ghosts. Together, they form a synthetic archive, a constellation of perspectives, an alternative table of contents for architecture today.
6.1 Poems
Here is the full list of 100 poems ON ARCHITECTURE,† and 3 to begin with:
6.1.1 Writing kills, drawing lives
on ARCHITECTURE * perspective: TIME ✼ Library: SerresFriends_Library_79
Writing has killed architecture. [1]
And yet the drawing keeps talking, time keeps the building alive as a joke that refuses to end.
To see space demands time, do not kill time. [2]
In the background is a piece of classical architecture. [3]
In the foreground, deadlines lean like scaffolds, comic and severe, the plan repeats the plan and calls it present.
Rome is rock; it is object; it has time; it makes time and requires time. [4]
Time, but time which has already rolle\d up, rounded itself, at the same time as it was splitting. [5]
We stack hours like floors, and floors like hours, until time itself becomes a hallway that turns back on its turn.
The architecture must find its justification in the deep intention of the Essay. [6]
[1] Serres, The Five Senses;
[2] Serres, The Five Senses;
[3] Foucault, Discipline and Punish;
[4] Serres, Rome;
[5] Deleuze, Cinema 2 The Time Image;
[6] Derrida, Of Grammatology.
6.1.2 Joke or Joule?
on ARCHITECTURE * perspective: ENERGY ✼ Library: Xenotheka_1523
Energy is information: information is energy. [1]
Information is information, not matter or energy. [2]
Energy is primary. [3]
And the building laughs softly: which is it, the joke or the joule.
New measures of architecture such as time, cost, or energy could be adjoined to length, width, and height. [4]
Fire and Memory: On Architecture and Energy. [5]
I stack rooms like arguments, push, pop, and repose; the plan performs its calories.
This is where the inconspicuous energy concept begins with the use of passive solar energy. [6]
We round the corners to keep the currents friendly; we square the meters to count the patience.
[1] Fuller, Synergetics;
[2] Wiener, Cybernetics or Control and Communication in the Animal and the Machine;
[3] West, Scale The Universal Laws of Growth;
[4] Witt, Formulations Architecture Mathematics Culture;
[5] Koolhaas, Elements of Architecture;
[6] Deplazes, Constructing Architecture;
6.1.3 Plan, measure, punchline
on ARCHITECTURE * perspective: MUSIC ✼ Library: Xenotheka_1523
Painting, sculpture, architecture, and all kindred arts do the same thing as music. [1]
We stack notes like bricks, we brick time like rhythm; the room hears us.
Unlike music, architecture is inescapably operative in human experience. [2]
I could finally build arches between music and architecture. [3]
So we hum in sections, elevations, a chorus of plans folding like measures.
’Music is architecture. [4]
Not only the architecture but also the physics can be designed at will—frozen music indeed. [5]
Joy is a load-bearing wall that laughs at gravity, timing its punchline with a pause.
Music is never tragic, music is joy. [6]
[1] Langer, Feeling and Form;
[2] Hays, Architecture Theory since 1968;
[3] Xenakis, Music and Architecture;
[4] Eshun, More Brilliant Than the Sun Adventures in Sonic Fiction;
[5] Houlgate, Hegel and the Arts;
[6] Deleuze Guattari, A Thousand Plateaus;
7 Meta-chapter: on writing this paper
This paper is not just about how architects might write with synthetic tools. It is itself written through those tools. What you are reading is not simply a study of a method—it is the enactment of that method.
The poems in this work—generated through a loop between Alice_ch3n81 and ChatGPT—explore what it means to write architecturally with algorithmic systems, biased libraries, and synthetic characters. But the paper you’re holding (or scrolling through) was written through a different loop: one between the author, ChatGPT, and specific data. This chapter documents that process.
Let us be precise.
The paper was initiated by the “author”—Dr. Miro Roman—who provided a detailed abstract, chapter titles, and a conceptual direction.† These outlined the structure and argumentative spine of the paper.
To support this foundation, four essential materials were shared with ChatGPT as database for its writing:
• The full article “1001 in 1: Recycling Stories at Light Speed”, which outlined a general method of working with ChatGPT in artistic and conceptual scenarios (Roman et al., 2023).†
• A transcript of the talk “
Day of Words” given on March 5th, 2024 at ETH Zurich, where the notion of living words, ChatGPT, Alice, language, and architecture was explored as part of a symposium (
Roman, 2024).
†• A transcript of the YouTube talk “
A Play Among Books”, a reflection on the book written with Alice, exploring its conceptual structure and computational strategy (
Roman, 2023b).
†• A transcript of the talk “
On Alice_ch3n81, GPT3, Midjourney, Botto and DALL·E”, where the distinction between different generative models and their aesthetic ecosystems was discussed (
Roman, 2023c).
†From this starting point, we proceeded in two loops of conversations with ChatGPT. First, we decomposed the source material, identifying key terms, arguments, metaphors, and methodological claims. These were then recomposed into new textual forms that followed the architecture of the abstract and proposed chapters. The first loop was generative—expanding each chapter into full-length sections of around 1000 words, complete with arguments, provocations, and transitions. The second loop was editorial—reviewing all chapters in light of the whole, adjusting voice, logic, rhythm, and scholarly positioning.
Throughout this process, ChatGPT served as a synthetic co-author, remixing citations, proposing formulations, and—crucially—responding to critique. Every section was read, edited, and refined through a dialogical exchange between human direction and machine fluency. The author asked for more references. He demanded less repetition. He removed romanticisms. He asked for precision. And the system adapted.
In this way, the paper mimics the very logic it describes in earlier chapters: a choreography of bias and synthesis, of input and flow. Except here, the actors are not Alice and GPT, but Miro and GPT. The method remains the same. What shifts is the context.
This is a key point: the paper is not just about using AI to write poems. It is also about using AI to write the discourse that frames those poems. And the same architectural concerns apply—how to select data, how to position one’s stance in a probabilistic field of possible sentences. The question is not whether one writes “by hand” or “by AI”—the question is: what mechanics do you construct, what data do you prioritize, and what argument do you want to make visible?
We make no claims that this process is neutral. On the contrary, it is biased by design. The stance was curated, and the materials were fed into ChatGPT with the explicit goal of sharpening—not softening—the position. ChatGPT is never used alone in this project; it always responds to something: a prompt, a set of quotes, a transcribed idea, or a curatorial logic. In doing so, this paper participates in what we have called architectonic writing: not a linear style, not an opinion, but a structured flow shaped by design. The entire process was designed—from the structuring of chapters to the quoting of talks, from the inclusion of footnotes to the construction of argument. Nothing was written “from scratch.” But everything was authored with care.
Is it too much?
Too fast?
Too synthetic?
Perhaps. But these are the very questions we must now ask about all intellectual production. Because the question is no longer whether you use ChatGPT. The question is what you do with it—how you shape it, challenge it, constrain it. In this moment of generative saturation, authorship is no longer about writing beautiful sentences. That part is handled. The boundary between human and machine is gone. From now on, it will always be mixed. The real task is not to police origins, but to sharpen arguments. Authorship is about selecting sources, defining loops, and constructing environments for writing and creating arguments.
This is what we did here.
A final transcript of the full conversation between author and machine is made available.† But in truth, the paper itself is the transcript. It is the trace of this dialogue, crystallized into academic form.
What architecture becomes—when written through a bouquet of intelligences—begins here.
8 Discussion
This paper has proposed and demonstrated a novel way of writing, organised as the synthesis of linguistic, computational, and coded operations mediated by artificial intelligence. The central claim—that architectural thought can be composed through the interaction of general AI, domain-specific AI, and human actors within a computational framework—was unfolded through the involvement of ChatGPT, Alice_ch3n81, and custom-coded API integration. We call this architectonic writing.
Chapters 1–3 introduced and differentiated the participating agents: ChatGPT as a generalized language generator, Alice_ch3n81 as a domain-specific knowledge interface, and their interrelation as a dynamic environment. Chapter 4 presented the methodology for producing 100 architectural poems—each generated through prompt logic shaped by human intention, biased databases, and model iteration. Chapter 5 contained the poems, and Chapter 6 reflected on the role of these very systems in writing the paper itself. This recursive structure underscored the central methodological hypothesis: that authorship and knowledge can now emerge from a choreography of human and machinic intelligences operating across new registers of scale, speed, and complexity.
These parameters are not merely computational byproducts; they are epistemic shifts. As
Hovestadt (2022) has suggested, architecture has long extended beyond buildings into the media systems through which it is communicated and performed. Computational writing continues this trajectory. The automation of architectural text at machine speed compels a reevaluation of editorial curation and authorial control. By accelerating cycles of composition and interpretation, AI systems enable new scales of architectural reasoning that are synthetic, relational, and iterative.
The integration of Alice_ch3n81—a custom search engine trained on architectural literature—was crucial to maintaining contextual grounding. Instead of treating bias as a flaw, Alice deployed selectivity as a conceptual device. In doing so, it reframed architectural research as projective rather than merely analytical. Alice served not only as a curatorial filter but as a structuring intelligence, steering the generative potential of ChatGPT toward architectural terrains. The resulting texts are not neutral outputs but filtered articulations—partial, contextual, and computationally lensed.
Similar approaches are already extending into built projects (
Plank and Roman, 2024), research (
Marincic, 2017;
Alvarez Marin, 2021;
Orozco Esquivel, 2017), and pedagogy (
Meteora, 2019;
0MORE, 2021), actively unfolding new ways how to approach architecture in the age of mass media. In Studio Meteora at ETH and 0MORE at UIBK, architectonic writing has become a key instrument for teaching and experimentation. The project Nifty Nook & Tashi Gfü exemplifies this trajectory: unfolding ideas through text, staging atmospheres through film, constructing narratives through images, and prototyping through models, while simultaneously engaging branding, social media, fundraising, logistics, and material treatments. Design here emerged as a synthetic continuum of thought, mediation, and construction. Within such a framework, the architect is less an autonomous author and more an orchestrator of tools, data, and structures of thought. Architecture today is as much about editing and curating informational environments as it is about form. Architectonic writing is not simply commentary on architecture—it constitutes a form of architecture, a framework for spatial, systemic, and synthetic exploration. The 1001 IN 1 (
Roman et al., 2023) project further demonstrated how this framework can expand beyond text to images and modeling, suggesting how design can operate across media in a world increasingly shaped by AI.
Finally, this research affirms that AI systems are not neutral vessels but active actors. They shape not only architectural conclusions but also the framing of architectural questions themselves. Architectonic writing thus becomes both method and medium, reflecting the shifting status of design in a world where art and science, logic and speculation, converge as entangled operations, much as the etymology of téchnē suggests. What is genuinely new is not the idea of synthetic authorship, but the scale, speed, and complexity with which it now unfolds. AI has captured language in ways unthinkable just five years ago (
OpenAI, 2025). Its performance is astonishing. The responsibility, however, rests with us: to decide how we will inhabit these instruments, how we will direct their power, and what kinds of authorship, reasoning, and design will arise from their entanglement with human intelligence.
2095-2635/2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.
PDF
(12413KB)
