The research investigates the relationship between architecture and generative artificial intelligence through both an experimental and a documentary approach. The reflection stems from a clear empirical premise: advances in the field of machine learning have introduced tools capable of producing coherent architectural representations by following principles and mechanisms that appear to mirror the way architects engage with the compositional process. In order to define a field of inquiry suitable for careful exploration and analysis, the research focuses on a specific form of generative artificial intelligence, namely latent diffusion models. The emergence of these tools is radically reshaping the role of the architect as designer, introducing an unprecedented form of authorship. Certain technical conditions, such as the use of images and the presence of probabilistic mechanisms, contribute to defining this new condition, which departs from the deterministic logic of computational and parametric algorithms. To highlight the relationship between architects and generative AI, the research introduces the concept of latent imitation as a theoretical device aimed at clarifying both the similarities and the differences between the two domains. The term imitation emphasizes how, in both cases, architectural representation derives from an imitative process grounded in image archives, while the attribute latent points to the specific nature of this method by focusing attention on the way representation is constructed within a latent space. The need to define a concept capable of describing this unprecedented way of constructing architectural representations constitutes one of the central aims of the research, a necessity that becomes implicit once the very definition of artificial intelligence is accepted. The attribute “artificial,” in fact, serves as an effective pretext for highlighting specific aspects of this technology, aspects that become even more evident when contrasted with the behavior of a hypothetical “synthetic” intelligence. From the comparison between “artificial” and “synthetic,” it is possible to infer that artificial intelligence imitates the outcomes of the human mind through profoundly different processes, whereas a synthetic intelligence would not only replicate the outcomes of human reasoning but would also reproduce its functioning. The broad range of applications introduced by generative AI is grounded precisely in this principle, namely the capacity to imitate the outcomes of the human mind through alternative paths and approaches. In architecture, artificial intelligence should therefore not be understood as a substitutive tool, but rather as an integrative element capable of enriching the compositional outcome through mechanisms fundamentally different from those employed by the “traditional” architect. Underlying the definition of latent imitation is the intention to emphasize the central role that datasets assume within the generative process. Datasets are not all alike, and their nature profoundly influences the type of tool built upon them. The role played by datasets, together with the probabilistic nature of the mechanisms governing generative AI, constitutes the foundational principle of this new form of authorship expressed through the notion of design responsibility. The concept of responsibility is itself deeply connected to the issue of datasets. Artificial intelligence tools are not all the same and, beyond distinctions based on technical differences, it is possible to establish an ontological classification according to the different types of datasets upon which these tools are built. Datasets may be closed, open, or collective, and changes in their nature also alter the nature of the tool itself and, consequently, the ways in which the designer exercises intention within the generative process. This condition raises urgent questions and demands a critical reassessment of the concept of design responsibility, a central issue within a context of hybrid human–machine production in which the boundaries of authorship remain uncertain and require a new theoretical and operational definition.
Il lavoro di ricerca indaga il rapporto tra architettura e intelligenza artificiale generativa attraverso un approccio sperimentale e documentale. La riflessione nasce da un chiaro presupposto empirico: i progressi nell’ambito del machine learning hanno introdotto strumenti in grado di produrre rappresentazioni architettoniche coerenti, seguendo principi e meccanismi che sembrano ricalcare il modo in cui gli architetti affrontano il processo compositivo. Per circoscrivere il raggio d'azione a un contesto idoneo a essere esplorato e analizzato con la giusta attenzione, la ricerca si focalizza su una specifica forma di intelligenza artificiale generativa, ovvero i modelli di diffusione latente. L'avvento di questi strumenti sta plasmando radicalmente il ruolo dell’architetto inteso come progettista, introducendo una forma inedita di autorialità. Determinate circostanze tecniche, come l’utilizzo di immagini e la presenza di meccanismi probabilistici, contribuiscono a definire questa nuova condizione che si dissocia dalla logica deterministica degli algoritmi computazionali e parametrici. Per mettere in evidenza il legame tra architetti e AI generative, la ricerca introduce il concetto di imitazione latente come dispositivo teorico per far luce su ciò che accomuna e distingue i due ambiti. Il termine imitazione sottolinea come, in entrambi i casi, la rappresentazione architettonica derivi da un processo imitativo basato su archivi di immagini; l’attributo latente indica invece la specificità di questo metodo, andando a porre l’attenzione sul modo in cui la rappresentazione viene costruita all’interno di uno spazio latente. La necessità di definire un concetto che qualifichi questo modo inedito di costruire rappresentazioni architettoniche costituisce uno degli obiettivi della ricerca, una necessità che diventa implicita nel momento in cui si accetta la definizione stessa di intelligenza artificiale. L’attributo “artificiale” infatti, si presenta come un ottimo pretesto per evidenziare alcuni aspetti specifici di questa tecnologia, che diventano ancora più evidenti se contrapposti al comportamento di un’ipotetica intelligenza “sintetica”. Dal confronto tra "artificiale" e “sintetico” possiamo dedurre come l’intelligenza artificiale sia qualcosa che imita gli esiti della mente umana attraverso processi profondamente diversi, al contrario, un’intelligenza sintetica non solo replicherebbe gli esiti del ragionamento umano, ma ne riprodurrebbe anche il funzionamento. La ricchezza applicativa introdotta dall’AI generativa si basa proprio su questo principio, ovvero sulla capacità di imitare gli esiti della mente umana seguendo percorsi e approcci diversi. In architettura, l’intelligenza artificiale non si presenta perciò come uno strumento sostitutivo, ma bensì come un elemento integrativo in grado di arricchire l’esito compositivo attraverso dei meccanismi ben diversi rispetto a quelli con cui lavora l’architetto “tradizionale”. Dietro la definizione di imitazione latente vi è l’intenzione di evidenziare la centralità che i dataset assumono all’interno del processo generativo. I dataset non sono tutti uguali e la loro natura influenza profondamente la tipologia di strumento che ne viene costruito sopra. Il ruolo ricoperto dai dataset e la natura probabilistica dei meccanismi che regolano il funzionamento dell’AI generativa, sono i principi cardine su cui si fonda questa nuova forma di autorialità espressa sotto forma di responsabilità progettuale. Il concetto di responsabilità è a sua volta profondamente legato al tema dei dataset. Gli strumenti di intelligenza artificiale non sono tutti uguali e, oltre a una distinzione basata sulle differenze tecniche, è possibile operare una classificazione di tipo ontologico sulle diverse tipologie di dataset con cui questi strumenti vengono costruiti. I dataset possono essere chiusi, aperti o collettivi, e al variare della loro tipologia cambia la natura dello strumento e, di conseguenza, cambiano anche le modalità con cui il progettista esercita le proprie intenzioni all'interno del processo generativo. Questa condizione solleva interrogativi urgenti e impone una valutazione critica del concetto di responsabilità progettuale, nodo centrale in un contesto di produzione ibrida uomo–macchina, in cui i confini dell’autorialità risultano incerti e necessitano di una nuova definizione teorica e operativa.
ARCHITETTURA E DATI. LA RESPONSABILITÀ DEL PROGETTISTA NELL’EPOCA DELL’INTELLIGENZA ARTIFICIALE Limiti e opportunità dell’utilizzo dell’AI generativa in architettura
GNASSI, FABIO
2026-05-27
Abstract
The research investigates the relationship between architecture and generative artificial intelligence through both an experimental and a documentary approach. The reflection stems from a clear empirical premise: advances in the field of machine learning have introduced tools capable of producing coherent architectural representations by following principles and mechanisms that appear to mirror the way architects engage with the compositional process. In order to define a field of inquiry suitable for careful exploration and analysis, the research focuses on a specific form of generative artificial intelligence, namely latent diffusion models. The emergence of these tools is radically reshaping the role of the architect as designer, introducing an unprecedented form of authorship. Certain technical conditions, such as the use of images and the presence of probabilistic mechanisms, contribute to defining this new condition, which departs from the deterministic logic of computational and parametric algorithms. To highlight the relationship between architects and generative AI, the research introduces the concept of latent imitation as a theoretical device aimed at clarifying both the similarities and the differences between the two domains. The term imitation emphasizes how, in both cases, architectural representation derives from an imitative process grounded in image archives, while the attribute latent points to the specific nature of this method by focusing attention on the way representation is constructed within a latent space. The need to define a concept capable of describing this unprecedented way of constructing architectural representations constitutes one of the central aims of the research, a necessity that becomes implicit once the very definition of artificial intelligence is accepted. The attribute “artificial,” in fact, serves as an effective pretext for highlighting specific aspects of this technology, aspects that become even more evident when contrasted with the behavior of a hypothetical “synthetic” intelligence. From the comparison between “artificial” and “synthetic,” it is possible to infer that artificial intelligence imitates the outcomes of the human mind through profoundly different processes, whereas a synthetic intelligence would not only replicate the outcomes of human reasoning but would also reproduce its functioning. The broad range of applications introduced by generative AI is grounded precisely in this principle, namely the capacity to imitate the outcomes of the human mind through alternative paths and approaches. In architecture, artificial intelligence should therefore not be understood as a substitutive tool, but rather as an integrative element capable of enriching the compositional outcome through mechanisms fundamentally different from those employed by the “traditional” architect. Underlying the definition of latent imitation is the intention to emphasize the central role that datasets assume within the generative process. Datasets are not all alike, and their nature profoundly influences the type of tool built upon them. The role played by datasets, together with the probabilistic nature of the mechanisms governing generative AI, constitutes the foundational principle of this new form of authorship expressed through the notion of design responsibility. The concept of responsibility is itself deeply connected to the issue of datasets. Artificial intelligence tools are not all the same and, beyond distinctions based on technical differences, it is possible to establish an ontological classification according to the different types of datasets upon which these tools are built. Datasets may be closed, open, or collective, and changes in their nature also alter the nature of the tool itself and, consequently, the ways in which the designer exercises intention within the generative process. This condition raises urgent questions and demands a critical reassessment of the concept of design responsibility, a central issue within a context of hybrid human–machine production in which the boundaries of authorship remain uncertain and require a new theoretical and operational definition.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



