Within the built environmental discourse, the potential applications of artificial intelligence tend to be discussed around a rather broad spectrum. At the two poles of the discursive range, there are competing conceptions of the nature and limits of artificial intelligence. At one pole, AI is conceptualized as a quantitative calculator operating at the behest of human-built-environmental professionals. On the other pole, AI is conceived as a fully capable synthetic competitor in emergence.

Part of this duality is shaped by whether the artificial intelligence system is ideally conceived around a black-boxed or grey-boxed framework (Chaillou, 2020-09). Within the former (black-boxed) structure, a human operator gives initial inputs to the synthetic operator, and the AI subsequently produces final outputs, with no interaction between human and synthetic occurring in between (ibid). Within the latter (grey-boxed) framework, the middle phase is opened up, allowing for more iterative and potentially reciprocal input-output cycles to take place (ibid). While the quantitative-calculator model could in theory be shaped around a black-box or grey-box approach, the synthetic competitor model must, given its discretion-based operations, be anchored around a grey-boxed approach at minimum.

As the discourse meanders towards the quantitative-calculator pole, the potentials of artificial intelligence are presumed to be best leveraged around the topics of: Energy and carbon footprints (Budennyy et al., 2022-12), environmental control systems (Park et al., 2020), structural systems (Salehi and Burgueño, 2018, 184), building codes and urban regulations [(Ilal & Günaydın, 2017); (Rafsanjani & Nabizadeh, 2023)], urban infrastructure [(Nikitas et al., 2020); (Zhang et al., 2021)]; post-occupancy analysis (Nayeri, 2023-06-20), and participatory design [(Llach, 2015); (Neuhauser & Kreps, 2011-03)]. On the other end of the spectrum, as the discourse meanders towards the synthetic-competitor pole, AI is seen to have discretionary and creative capacities in addition to the quantitative areas noted above, allowing it to be leveraged within the iterative or conceptual design process itself [(Barker, 2023-04-26); (Reinholdt, 2023-06-03); (UNStudio, 2023)].

Regardless of the discursive location within which it is placed, AI is clearly understood as a new and lasting actor within the built environmental landscape. At a minimum, it is seen as having the potential for increasing efficiency and productivity within built environmental authorship. This increased capacity could on the one hand allow smaller firms to compete better with larger firms, or it may pull the momentum further towards the side of larger firms who may be able to more readily take advantage of new technological developments (Leach, 2023-02-13). At the more disruptive end, AI is presumed to be on a path that will lead to its independent entry into the market of built environmental authorship, engaging in direct competition with human actors—wherein it is oft presumed that the free market competition between synthetic and human will favour the former, and make obsolete the latter (Leach, 2023-02-13).

In contrast to the bipolar spectrum noted above, the discourse concerning the leveraging of AI at the scale of urban infrastructure leans predominantly, if not ubiquitously, towards the potential of AI as a quantitative prosthetic. To give some examples from recent pieces occupying representative niches within the contemporary discursive landscape: (Lutfiani et al., 2024) focuses on AI’s potential for the optimization of resources and energy use, and for the creation of more efficient infrastructural monitoring and control systems; (He & Chen, 2024) focuses on AI’s capacity for “predictive analytics, decision-making improvements, and the automation of complex geospatial tasks in urban areas;” offering a broad survey of the urban planning profession within the United States, (Sanchez et al., 2023) asserts that much of the profession considers the most straightforward potential of AI to lie in the realms of demographic and transportation-network analysis; (Forkan et al., 2024) and (Rezvani et al., 2024) note the potentials of AI as a “proactive detection” tool for transportation networks; (Sepehr, 2024) focuses on a comparable quantitative application of AI within traffic light systems; (Jagatheesaperumal et al., 2024) tackle similar quantitative transportation network issues except via the slightly-more specific lens of smart city frameworks; and (Dai et al., 2024) take this capacity for AI to monitor critical predictors to the domain of urban ecological systems.

In general, there is a marked paucity within the discourse of urban infrastructure of texts noting the more qualitative potentials of AI. It is only when the scale of concern shifts down to the actual design of urban form, landscape architecture, architecture, or interior architecture that references to AI with regard to its more creative or qualitative potentials begin to proliferate.

In its current form, the most problematic aspect of ChatGPT is its fundamental inability to confess a lack of knowledge. “I don´t know” is simply not within its programmed vocabulary, leaving the platform to generate farcical information or data points, particularly when faced with limited or inadequate datasets, that can be misleading. (Alier Forment et al., 2024) refer to these as “hallucinations.” A rather egregious form of this is when ChatGPT is asked for a list of articles relevant to a subject, to which it often generates a list of non-existent articles in non-existent journals.

Within generative AI platforms that produce visual media, such hallucinations take on a slightly different tone. For instance, a simple prompt of “a group of people holding their hands in front of their faces” can produce what seems like a fairly accurate image; however upon further inspection, various physical anomalies often emerge—in this case, a number of six-fingered hands, unusually shaped thumbs, appendages which seem to emerge from and belong to multiple persons, and so forth (see Figure 3).

Throughout the studio, once this capacity for AI platforms to produce hallucinations became apparent to students, not as an anomalous, but rather as a consistent tendency, a range of different outcomes began to emerge.

With regard to ChatGPT, students no longer regarded the platform as a source of infinite and infallible knowledge, but rather as an aged uncle (as per a student’s description) that had accrued significant amounts of knowledge, but also was prone to going off on fictional tangents. As a consequence, students began to engage the platform as a supplementary, not primary, research tool. Questions to ChatGPT were limited to more strategic or conceptual approaches to blue-green-gray infrastructure and net-positive urban energy systems. The instructor of the course, along with peer-reviewed publications, functioned not only to open up deeper areas of investigation but also as a baseline fact checker for ChatGPT´s initial analyses.

In general, ChatGPT performed accurately when framing the overarching concepts, strategies, and ideas in use within each discursive domain. However, when delving into more specific and fine-grained areas, there was an increased likelihood of the platform generating fallacies. With regard to Midjourney and other image-generating AI platforms, such hallucinations proved to be productive in different ways.

Each group was required to produce a set of five large-scale background images, via image-generating AI platforms, focusing on different built environmental scales. The background images were initially prompted to accommodate a series of systemic interventions as framed by the students. However, the produced images often had errors or noise of various kinds—e.g., infrastructural systems that illogically overlapped with buildings and built-environmental conditions that collided in manners that betrayed the basic fundamentals of construction systems (or physics). Student discretion and knowledge were subsequently required to refine these elements back into a workable order.

Through this process, new built environmental ideas often emerged. Per the students’ own acknowledgement, the outcome was often a built world condition that was more “layered” than what they would have achieved if the AI platform had not been part of the process. This is in line with the findings of (Liu, 2024), which notes that “errors, while challenging, also serve as unintended prompts for students to consider alternative designs and solutions, fostering a space for innovative thinking.”

Of critical import here, is that AI-imagery was not unidirectionally molded around human-authored project narratives. Rather, AI platforms were invariably granted a partial co-authorship capacity and were allowed to reshape and refine the initial narrative set forth by the human actor. Whether or not students consciously recognized or acknowledged said co-authorship however, is a different subject altogether, as discussed in Section 3.3.

ChatGPT´s utility as a sounding board was much more consistent when compared to the above-mentioned areas of use. Groups could ask, “We are doing x, y, z with this proposal, but we are also trying to do a, b, c—is there a logical discrepancy across these efforts?” To such questions, ChatGPT proved quite capable of underscoring anomalies or contradictions to the overarching project narrative and suggesting alternative pathways for consideration (see Figure 4). This is in line with AI´s theorized potential as a type of “thought processor” as discussed by prior scholars within the artificial intelligence field (Kolata, 1996-12-10). What is of particular intrigue is that the utilization of ChatGPT as this manner of thought processor allowed for the emergence of a third party within the design studio culture.

In this new role, ChatGPT was utilized in two manners. First, as an internal logic checker which students could deploy individually, to verify if their own narratives concerning a specific area of a project were indeed sound. Second, as a mediating or arbitrating voice in discussions.

The architecture studio is, at least in theory, a place of exchange and collaboration. However, structures of hierarchy, whether real or perceived, may inhibit this ideal, and the number of voices offering different perspectives can become limited. The worldviews, ideologies, and inclinations of the instructor may be inadvertently mirrored by some students. In some cases, these frameworks may be imposed, consciously or unconsciously, by the instructor upon the students in a hierarchical manner. Groupthink dynamics may create artificially cohesive clusters within the student cohort. There may also be a real or perceived hierarchy amidst the student body based on prior academic or creative performance.

It is quite rare for a student to be able to receive external feedback (whether from an instructor or a peer) and not have these structural hierarchies, biases, issues of internal competition, and peer pressure, inhibit the depth to which said feedback is received. ChatGPT is quite an anomaly in this regard, in that it clearly functions outside of the studio culture, and outside of its real or perceived dynamics.

Within the course, particularly when dialogues internal to a group, or between a group and the instructor, became locked in place around a bi-faceted structure, ChatGPT was brought in to help disentangle the polarized state of the discussion. Initially, the deployment of ChatGPT in this manner was coincidental. However as the studio progressed, this became a pedagogical tool that was deliberately deployed towards this end. This role as a disentangler was also aided by the perception of ChatGPT as a more neutral participant without a stake in the outcome of the overarching project.

For the instructor, ChatGPT thus served as an easily accessible second point of view which could help underscore the vulnerabilities of a position that a student was defensively holding onto. For the student, the platform served as a readily accessible second point of authority that could serve to counter or verify the instructor’s argument. This was particularly useful in the context of a studio which can oftentimes fall to, or at least be perceived as falling to, a clear hierarchical structure, laced with internal structural biases.

Both at the start and near the terminus of the course, students were queried about how they understood the nature of authorship in the context of the visual media generated via AI platforms. This was structured as an informal discussion which the instructor subsequently recorded the details of in the form of field notes.

At the beginning of the course, 10 out of the 16 enrolled students asserted that they were undecided on the matter of authorship, and had to further engage with such image-generating AI platforms to better understand the situation. Four asserted that ownership of the content produced belonged to the AI. One student asserted that the authorship belonged to the human user inputting the prompts, while another asserted that there was some manner of human-synthetic co-authorship taking place.

By the end of the course, 13 leaned towards the position that there was some form of human-synthetic co-authorship taking place, with varying degrees of collaboration. Two students asserted that the imagery produced was purely the intellectual property of the generative AI platform. One student asserted that the imagery was entirely the intellectual property of the human user.

Across the course, students also used artificial intelligence platforms to periodically mine for imagery to help move past creative blocks. Students asserted that this was similar to the process of mining across image-archiving platforms such as Pinterest (i.e., Pinterest mining). However, students felt much more liberated to intellectually or creatively absorb the images generated via AI platforms, since their perceived synthetic origins allowed students to avoid the fear of replicating something that had already been done by another human hand. Even students who stressed that AI-generated imagery was the intellectual property of the AI platform confirmed this inclination—the main reasoning being that said imagery was not the product of a prolonged effort by the AI platform, but rather something produced with great rapidity, as if in passing.

Excerpts from the final deliverables of three of the four group projects generated via the studio are shown here. The first (Energetic Madrid) deals with urban energy systems, and the latter two (Flood Harvest and Waterkeepers) are focused on blue-green infrastructural systems. For brevity, the fourth project generated via the studio, focused on urban energy systems, was omitted from the paper.

Energetic Madrid, by Ghofrane Benajiba, Daniela Figueroa, Nathan Robert, and Tristan Sartorius, focused on the scale of the street, building (Figure 5), block (Figure 6), plaza (Figure 7), and superblock to establish a net-positive-energy infrastructural system throughout the city. Solar heat chimneys, trombe walls, night-flush cooling, biomass-based heat generation, seasonal thermal storage systems, photovoltaic energy generation, reductions of domestic consumption footprints, heat recovery from industrial programs, and fifteen-minute city models were some of the primary elements deployed throughout the work. The end outcome was a mixed-use urban condition with: (1) reduced carbon and energy footprints, (2) heightened energy generation capabilities (achieving cross-scalar energy surpluses subsequently stored in the form of heat within seasonal thermal energy storage systems), (3) and in-built proximities to critical urban resources modelled off of fifteen-minute city ideals.

F(l)ood Harvest, by Alara Aykanat, Yusuf Bozkurt, Daphné Fournel, and Martin Jitenski focused on the domestic scale (Figure 8), the scale of the street (Figure 9), superblock, neighborhood (Figure 10), and city scale to establish a two-faceted water infrastructure capable of managing severe drought and flooding conditions. Set 150 years into the future wherein the extreme scenarios of predictive models for flooding and water scarcity were taken into account, this project was an extensive exploration of the sociocultural, political, economic, as well as infrastructural dynamics that could emerge in such circumstances (Figure 11). Blue-green roofs, water retention and detention systems, rainwater harvesting, low-water plumbing fixtures, wind scoops, localized food production and storage, bioswales, water plazas, localized blackwater/greywater remediation, autochthonous landscaping, permeable hardscaping, multimodal (and interspecies) street design, and revitalized wildlife and riparian corridors were some of the elements put in play within the project narrative.

Waterkeepers, by Nadine Fahmy, Malek Fleifel, Roché Rabie, and Carolina Schönburg, homed in upon the scale of the building (Figure 12), boulevard (Figure 13), block, plaza, and city (Figure 14) to establish a flood and drought tolerant water infrastructural system for Madrid while also overlapping with fifteen-minute city ideals. The main systemic elements utilized include green roofs and facades, localized water retention and detention tanks, decentralized greywater/blackwater remediation systems, rainwater harvesting, bioswales, permeable hardscaping, water plazas, soil-based carbon sequestration, urban timber production, and revitalized riparian and wildlife corridors. The model of the city that emerged was one that combined nature-based stormwater and drought management systems and mixed-use fifteen-minute city ideals, with significant capacity for soil-based carbon sequestration.

In the context of these three sets of deliverables, three outcomes were observed from the use of AI, particularly when compared to the prior year which did not lean on artificial intelligence in any format. First, the overall quality and detail of the final deliverables were at a significantly higher level when compared to the prior year’s course. Second, despite varying stances among students regarding the intellectual ownership of AI-generated content, there was no observable difference, when compared to the prior year, in the students’ sense of ownership or authorship over the final executed work. Third, perhaps contrapuntally to the prior point, AI clearly behaved in a coauthorship capacity throughout the iterative design process—not as a black-boxed computational prosthetic, but rather as a grey-boxed structure with which the students were engaging in active dialogue, even if without explicit recognition.

Particularly intriguing, was that the image generation process created a dynamic that the students could not control in full. AI-produced visuals were always embedded with unexpected urban spaces, infrastructural conditions, aesthetic parameters, unrealistic built environmental collisions, etc., that lay beyond the boundaries of the initial prompts. These anomalies in turn tended to alter or inform existing conceptual paths, narratives, and strategies that had been set in motion by the student groups. As the students themselves noted, this back-and-forth interaction between human and synthetic author, pushed the final deliverables to be more layered and more intricate than what would have been achieved if the AI platforms had not been embedded within the iterative design process.

There were no new infrastructural ideas that emerged from the use of AI in this manner. However, there were unexpected overlaps between infrastructure and urban life that arose through the process. Across the breadth of the deliverables seen above, water and energy systems are leveraged in ways that directly impact the physical fabric of the city in a manner that is deeply tied to human (and non-human) life. Very little of the infrastructure is hidden from view. This is a particular point of intrigue within the discursive landscape of urban infrastructure, wherein the visibility and sociocultural (and by extension socioeconomic) impacts of infrastructure are often a critical hurdle in terms of its public acceptance—with blue-green-gray infrastructure being one of the few exceptions to this (Toth, 2020-09-02).

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sector/ individuals.

Not applicable.

The author declares that there is no competing interest.