H2M’s Emily Fusilero and Robert Nilsen Were Recently Featured in AIA NYS’s October 2023 Issue Discussing The Responsible Use of AI in Architecture

H2M’s Emily Fusilero and Robert Nilsen Were Recently Featured in AIA NYS’s October 2023 Issue Discussing The Responsible Use of AI in Architecture

Delve into the intricate world of artificial intelligence in Emily Fusilero’s and Rob Nilsen’s latest essay, exploring how human innovation shapes AI while confronting its imperfections and inherent biases. Uncover the critical importance of understanding AI’s origins, training data, and biases to wield its power responsibly, unlocking the key to true impartiality and groundbreaking advancements. Read more below:

Training a Panacea: Responsible Use of AI in Architecture

Artificial Intelligence (AI) stands as a testament to human ingenuity, yet like other tools sculpted by human hands, it’s not devoid of the imperfections and biases inherent to its creators. It’s essential to dissect and understand the very fabric of AI tools — the data and training they receive. Situated within the architecture and design disciplines, how such tools can be used to exacerbate or alleviate issues surrounding the built environment. Who crafts these sophisticated models, and from where does their training content originate? While they promise groundbreaking capabilities, these tools are only as unbiased as the data fed into them. For architects and designers to harness the transformative potential of AI responsibly amidst a multitude of controversies, one must approach it with a three-pronged panacea: a thorough understanding of its inner workings, a vigilant awareness of its biases, and a constant strive for impartiality.

While mainstream media often presents AI through a sensationalized lens, technologists and academics have underscored the urgency of fostering nuanced dialogue about its broader ramifications and potentialities.1 The emergence of AI-based tools in architecture further necessitates such conversation. Though our intent is not to present novel findings or prophesize on future trajectories, we seek to root our discourse in the present context. Acknowledging that design tools for the built environment have perpetuated disparities and disenfranchised certain communities, we aspire to queer their appreciation amid AI’s early uses.

In the realm of architecture and design, the artificial intelligence mystique can be clarified by focusing on unsupervised machine learning and generative AI, which form the backbone of many tools in the industry.2 3 Unsupervised learning can be best understood through a practical architectural example: Studying a diverse collection of building designs without labels or styles, and being tasked to identify any inherent patterns within the geometry. Analogously, unsupervised learning models uncover structure in data without any explicit labels, clustering them by seeming relation. The application and comparison process is mathematically grounded with optimization techniques: Loss functions and the like, to identify relationships between the data. By utilizing this training data and incorporating generative properties and models like GANs or generative adversarial networks, can produce new forms of data…that is, in the context of our previous example, generative models could create new building designs based on the patterns or themes they’ve learned.

Within the domain of generative models, a key principle involves the transformation of training data into latent representations. These high-level representations serve as compacted, abstracted encodings that encapsulate the fundamental characteristics of the data. Analogous to architectural design where foundational themes or motifs are distilled from a myriad of influences, generative models systematically categorize and encode salient patterns from their training data. The significance of this latent space becomes particularly pronounced when one examines the potential for interpolation within it. By methodically navigating and interpolating between discrete points in this space, one can derive transitional designs that synthesize distinct elements, thereby revealing innovative configurations. For example, in the context of creating a façade generator, this would emulate a stepped transition of a façade with windows to no windows, or from glass materiality to stone. This process offers a computational parallel to the architectural practice of merging foundational principles from diverse styles to conceive a singular, cohesive design, all underpinned by the model’s synthesized comprehension of the inherent data patterns.

Ergo, the significance of these applications for creativity within the architectural design industry is clear. These models allow us to augment our conception of design by providing a novel means of conceptualizing similarity or difference.4 In addition, the introduction of generative properties to reproduce these conceptions enables a plethora of opportunities for designers beyond the built environment. Yet, it is imperative to acknowledge that these applications are tools, and as such, are indistinguishable from the inherent dispositions of their creators. Although enshrouded by complexity and frameworks distant from common understanding, these tools require a level of scrutiny and criticality akin to that demanded by historical technological advancements and innovations.

As discussed by professor and academic researcher, Joanna Zylinska, “The roots of AI can be traced back to Aristotle’s work on formal logic: more specifically, the syllogism, i.e. a form of deductive reasoning that allows one to reach a conclusion derived from a number of statements with an assumed truth-value.”5 This relates to the explicit responsibility of humans’ critical thought in developing the rules and training of a ‘non-human’ mode of machine learning — that is, the framework of an AI model’s agency to think and produce. Taking the CycleGAN as an illustrative case, when specific training elements are employed to generate a novel image, the resultant comprehension of textures, rules, and inherent biases are fundamentally anchored to the input dataset.6 Put succinctly, the artifact reflects the predispositions of its architect. Not unlike other technologies, AI has the ability to propagate such biases.

(Pictured L-R: Real Image A; Real Image B; Fake Image B; Fake Image A) 

In utilizing AI as an instrument for augmenting or reshaping the built environment, it becomes indispensable to grapple with the historical and sociopolitical dimensions of architectural practice. Elements of architecture have historically manifested and perpetuated systemic biases. The concept of disentanglement, wherein intertwined factors of variation in the data are separated into distinct and interpretable variables, offers a promising avenue to examine and mitigate these biases. By achieving a high degree of disentanglement, one can isolate and inspect specific biases or patterns, offering both a clearer understanding of the data’s underlying structures and an effective means to address and mitigate any unwarranted biases present. In combination, when integrating into machine learning applications, careful consideration of training material can alleviate issues of perpetuating systemic prejudices. Therefore, these elements of architectural history must be responsibly utilized as references to avoid the unintentional implementation of these discriminatory elements. Uninformed incorporation risks echoing long-standing forms of oppression.

By combining the power of AI with a holistic education of the built environment, architects have the responsibility and ability for xeno-solidarity — to extend empathy and consideration for other — a concept introduced in a new wave of twenty-first century feminism: Xenofeminism. As discussed in Xenofeminism: A Politics for Alienation, “Serious risks are built into these tools; they are prone to imbalance, abuse, and exploitation of the weak. Rather than pretending to risk nothing, XF advocates the necessary assembly of techno-political interfaces responsive to these risks.”7 8 Awareness and responsible utilization of these tools by architects and designers can aid in our process for a more productive, unbiased, and accessible future in the built environment that is thoughtfully inclusive to all identities in our society, regardless of race, class, gender, or ability.

Increasingly, designers must operate simultaneously at multiple scales (such as the urban, architecture and the built environment, objects, things and bodies) and often contradictory perspectives (including human as well as nonhuman stakeholders) — to remake the collaborative, peer-produced, open-source city.9 Our responsibility as designers in the twenty-first century has now moved beyond the necessity for program, form, and function and has extended into the realm of the politics of identity and inclusivity as a result of research, advocacy, and acceptance of diversity within the discipline.

The collaboration between AI and architectural design presents both unprecedented opportunities with inherent risks and challenges. The use of AI technology in the future of the built environment requires an ethical grounding of commitment to inclusivity, understanding, and historical context as aligned with xeno-solidarity and new wave Xenofeminist theory. For who and what does AI benefit? What positive and negative elements of human intervention can it discern? In the words of Joanna Zylinska, “Who and what can’t it see?”

Click here read the full Architecture New York State October 2023 issue from AIA New York State.

[1] For mainstream media examples, see “‘I am, in fact, a person’: can artificial intelligence ever be sentient?” by Amelia Tait at The Guardian, 2022, or “Conversing with a self-aware AI” by Craig Thomler on Linkedin News, 2021, or “No, Artificial Intelligence Isn’t “Alive”…Yet. But It Will Be” by Sam Westwood on Icon, undated. For examples within the ML / AI discourse, see “Sparks of Artificial General Intelligence: Early experiments with GPT-4” in arXiv:2303.12712, Cornell University, by Bubeck et. al, 2023.

[2] For example, see “Getfloorplan” (https://getfloorplan.com/), an AI tool that generates detailed 2D and 3D floor plans with 360 degree virtual tours, and “AI Room Planner” (https://airoomplanner.com/), an application that generates interior design concepts with staged furniture and different room styles, and “ARCHITEChTURES” (https://architechtures.com/en/), a generative AI-powered building design platform.

[3] Image-to-image technology (unsupervised learning) is used to allow user engagement for a VR Art Museum, built with Stable Diffusion Image-to-Image pipeline. Harvard Graduate School of Design, by Robert Nilsen, 2022-23.

[4] Carter and Nielsen, Using Artificial Intelligence to Augment Human Intelligence, 2017. Proposed that new devices for creativity should augment human ability to create.

[5] Joanna Zylinska, AI Art: Machine Visions and Warped Dreams, 2020.

[6]  CycleGAN, a high-level generative adversarial network, is a machine learning tool to create simulated images. The final output from this computational training was the result of selected images from the new terrain, generated by the texture mapping of rhizomorphic mycelium, and the satellite image of a selected site in East London. The new terrain informed the cycleGAN machine learning tool of areas that are textured for urban development, shadows for greenspace, and a flat surface for the formation of water bodies. The site image provided was a source for the cycleGAN to understand the graphic representation of urban density, green areas, and waterbodies. Selected training images from these two sources were carefully curated. After training through hundreds of epochs, the final image was realized. The Bartlett School of Architecture UCL, by Emly Fusilero, 2022.

[7]  Solidarity between individuals can be made to be more inclusive, xeno-solidarity, a philosophical theory proposed by Helen Hester. She addresses the ways that the binary and heteronormative collective thinking can further hinder the progression of the individual’s ability to extend what she refers to as “care and hospitality” to those whose identities we may deem to be considered “other”.

[8] Xenofeminism seeks to confront the responsibilities as a collective agent capable of moving between different levels of political, material, and conceptual organization. Laboria Cuboniks, theorists of Xenofeminism elaborates further that “Xenofeminism constructs a feminism adapted to these realities: a future in which the realization of gender justice and feminist emancipation contribute to a universalist politics assembled from the needs of every human, cutting across race, ability, economic standing, and geographical position.” Intervention in materiality is equally as important as intervention in digital and cultural. Laboria Cuboniks, Xenofeminism: A Politics for Alienation, 2018, 3.

[9] Forlano connects these theories and the networking of hybrid entities, issues, and infrastructures that make up the urban fabric and happenings of a city. Rather, designers can operate as mediators of these forces and advocate for less visible nonhuman stakeholders. Cities worldwide are currently rushing to build sensor networks capable of tracking systems and human behavior. Laura Forlano, Decentering the Human in the Design of Collaborative Cities, 2016, 48.