• “For the moment, I see [AI-generated images] very similar to sketching,” says Andrew Kudless, founder of Matsys. Kudless’ AI “sketch” perhaps recalls a veiled version of Marlon Blackwell’s Keenan TowerHouse (Fayetteville, Arkansas, 2000). As Kudless reminds us, “Design is about so much more than sketching; it’s more than coming up with an idea.” AI-generated images are not particularly useful as resolved design concepts. Instead, as collections, they start to elucidate a feeling, a vibe, a possibility. - image by Andrew Kudless

In this digital era of custom-tailored algorithms and hyper-personalized feeds, it’s hard to know if any given person is actually seeing the same things happening in the world as anyone else. Yet it’s no stretch to suggest you’ve probably noticed an explosion of artificial intelligence (AI) in the popular imagination over the last few months. Last year, web-based AI image generators DALL-E and Midjourney emerged on the scene, and ChatGPT, a text generator from OpenAI, made waves soon after. More recently, you may have heard about a New York Times reporter who had a disturbingly existential conversation with Bing’s AI searchbot, or of prominent art critic Jerry Saltz’s cynical takedown of Refik Anadol’s “Unsupervised,” an AI-generated painting in the main lobby at MoMA, as a “glorified lava lamp.” No doubt several more headlines will have been made in the short time between the writing and publication of this very piece.

It’s official: AI is having a moment. But what exactly is that moment, and what does it mean for architects? Is it the beginning of a new aesthetic or technological movement? A faddish flash-in-the-pan? Is AI coming for all our jobs? Is it simply a new tool on our proverbial belt, ushering ever-increasing efficiencies into the minutiae of our labor? Or could it bring about a new world entirely?

For now, at the presumed onset, the jury’s still out. According to Saltz, AI-generated art is “derivative and familiar and struggles to transcend its source material.” He says, “If AI is to create meaningful art, it will have to provide its own vision and vocabulary, its own sense of space, color, and form.” Though a biting critique, this is a more nuanced take than the ones that have been circulating the web after artists understandably made an uproar about AI image generators using their work as training data without their knowledge or consent. From this perspective, AI is simply regurgitative and exploitative. If we follow Saltz’s criticism, however, it is assumed that there could be such a thing as meaningful AI-produced art (and architecture), but for that to happen, AI would need to be a truly generative, even revolutionary, tool. It would have to produce something new or disturbing or unknown, rather than simply remixing and sampling previous hits (which AI is, admittedly, very good at). But the unknown is a risk, and as AI is a product of Silicon Valley, one can already see indications of a technology being rolled out for market adoption as soon as possible without the ramifications of that technology being fully considered: The winners profit; consequences be damned. 

Dr. Kenneth R. Fleischmann, professor and director of undergraduate studies at the University of Texas at Austin School of Information, is a leading expert on ethics in AI. When asked about the risks of the tech industry’s attitude of growth before everything, he points to Facebook as an example of an ubiquitous technology gone awry. “Based upon what we have learned about Facebook from whistleblower Frances Haugen … the issue is about not only making these tools as widespread as possible, but designing them in ways that maximize engagement at all cost, particularly in service of ad revenue,” says Fleischmann. “Relying entirely on the private sector to self-police themselves is particularly fraught.” For Fleischmann, a computer scientist, the issue comes down to designing good systems — a relatable refrain for architects. To paraphrase technology theorist Benjamin Bratton, AI is a human construct which has not yet been fully determined; AI will be what we make AI to be. Says Fleischmann: “Depending on how AI is designed, developed, and deployed, the increasing adoption of AI may lead to a more equitable society where AI is used to do the most good for people with the most need, or to a more dystopian society where AI leads to a greater degree of wealth inequality and reduced human freedom.”

If AI is to be a truly revolutionary tool for our labor and our world, we must resist its commodification toward maintaining the world as it is. As we’re already seeing in the “chat wars” between Microsoft, Google, and other massive tech players, it’s already being commodified, and the status quo holds the simple advantage of inertia. But architecture — from the initial concept sketch to the tedium of construction administration — is predicated on envisioning, communicating, and bringing into being a world that is previously unseen or unconsidered. For this reason, architects are particularly equipped to engage with AI tools, not only because we have become accustomed to using software in our design process, but because of our critical engagement and curiosity about the ways things are and how we live. For the architect, AI holds the potential of being not simply another tool or process, but an entirely new state of living.

Political philosopher Jacques Rancière noted that political failures are most often due not to a lack of awareness of the current issues at stake by the average citizen, but rather to a failure of imagination to create a compelling alternative to the status quo. The devil you know, of course, is better than the one you don’t know. Cynicism is an appropriate response upon hearing that Microsoft is investing $10 billion in its ChatGPT research while simultaneously laying off 10,000 employees. But according to Rancière and journalist Aaron Bastani, author of “Fully Automated Luxury Communism,” it is hopefulness and an expectation of abundance, not just criticality, that are needed. Bastani’s 2019 book lays out an optimistic vision for a post-work, post-capitalist, pro-technology, AI-assisted future on a sustainable and equitable earth. It is, perhaps, overly optimistic to the point of accelerationist naivete, but it at least paints a picture that we can debate of what the world could be — just as important architecture can do. 

But what is it about AI that makes such visions of grandeur worth discussing now? Though it is born out of Silicon Valley and is tainted by a dark skepticism in popular culture that dates back at least to HAL 9000 in Kubrick’s film “2001: A Space Odyssey,” one thing AI is incredibly good at is accessibility. Andrew Kudless, the founder of Matsys and the Bill Kendall Memorial Endowed Professor at the University of Houston’s Hines College of Architecture and Design, discussed the accessibility of and his work with text-to-image (TTI) generators like Midjourney and Stable Diffusion. If you’ve come across AI-generated architecture on the internet in the past year, you’ve almost certainly seen Kudless’ work. As of this writing, Kudless has accumulated more than 100,000 followers on Instagram (@matsysdesign), and his most viral posts have topped 175,000 likes. 

Accessible Design

For Kudless, the opportunity that TTI tools afford is a much lower barrier to entry, particularly for students, than traditional design tools like CAD, BIM, or 3-D-modeling software. But it’s less about instantly creating highly polished images than about richly conveying an idea quickly without the prerequisite training in software or architectural drawing. Kudless explains: “For the moment, at least, I see [AI-generated images] very similar to sketching. I think there’s somewhat of a misunderstanding because they look like renderings, so we tend to think of them as renderings. But the way renderings operate for a designer is they tend to be something that we don’t design through all that much, because of how laborious they are to create. AI images are almost the opposite.” 

Even for an accomplished computational designer like Kudless, TTI generators, in particular, have made the process of working with AI so much simpler than working with a propriety model, which was previously required. Says Kudless: “Even though I can write things in Python and have experience doing code, like a lot of people, I’m self-taught and make a lot of mistakes. I [thought], okay this is not accessible enough for me yet…. It’s something a lot of architects struggle with. Not just AI, but Grasshopper, Rhino, Revit — they all have varying degrees of accessibility, and it often takes a long time to become fluent in any of these tools. And I just wasn’t making much progress over the last couple of years until, all of a sudden, [I’m] in a little private chatbot with Midjourney. It’s completely turned the table, in a way.” 

Of course, an instantly generated collection of random pixels that an AI model determines “yes, this is what you’ve asked me for” and that has no inherent geometry does not a resolved work of architecture make. This imagery often takes on an Escherian or even Cubist quality, with multiple vanishing points and impossible geometries. There is (currently) no intelligence behind the objects shown except for the pixel on the screen — no underlying mesh object, as is the case with traditional rendering engines. But, as Kudless reminds us, that’s not the point. They are not particularly useful as resolved design concepts. Instead, as collections, they start to elucidate a feel, a vibe, a possibility.

Accessible Audiences

Though it is easy enough for architects who work with sketches, models, and renderings to understand that different media are useful for conveying different scales, levels of detail, and content, those who do not have architectural training may not catch on to those nuances. One of the more interesting conundrums of the explosion of AI imagery is that it, of course, has not been exclusive to architects and designers (as a new release of Grasshopper might be) but instead has had a much broader reach in popular culture. As such, images of fantastical, artificial architectures have been shared much more widely on social media than our intradisciplinary drawings typically are. 

As Kudless’ images went viral on Instagram, he noticed this shift. He says: “I already had a pretty large following of architects and designers, and that’s my core audience. It’s my peer group who I want to hear from. So when that group changed its composition, it was really strange for me. I began to realize that my new followers had a very different understanding of the work. They weren’t as critical, let’s say. They were people who wanted to have weddings or photoshoots; they don’t see the work the way I see it. That was another insight to me. I wasn’t really speaking to the same audience that I had before.” 

And although this new audience may not have been as critical or perceptive (“this isn’t a real building nor are we claiming that it should be”), with this expanded scope comes opportunity. Kudless explains: “I’m excited about how accessible it is in relation to working with clients.… Now we have a tool that allows clients to speak in the language of images very quickly and to produce images which appear, at least on the outside, very sophisticated when in fact they’re not. They are kind of cheap, in a way. But it opens up an opportunity for architects, because it allows us to collaborate in a new way with the client.” And while some people worry that this accessibility for the untrained could put the future of the architect at risk, he disagrees. “Some people see this as an existential threat to the profession,” Kudless says. “I don’t see that at all. Design is about so much more than sketching; it’s more than coming up with an idea. There’s a tremendous amount of time and labor that it takes to translate a sketch into an actual building.”

Accessible Data

Though they’ve expanded our potential audiences and shortened the time required to get to an interesting image — going back to Saltz’s criticism of Anadol’s work — there’s a question about what exactly those images are “creating” and what they’re drawing from. Models like DALL-E 2 from OpenAI, the same company behind ChatGPT and Bing’s searchbot, are completely closed and proprietary. No one in the general public knows what’s going into the dataset it’s trained on, and user data is fed back into the model to retrain it. Midjourney isn’t open-source either, but it is more transparent at least. But there’s a strange feature to its model: Every image that goes into the training data that the model draws upon when given a prompt has been scored on its aesthetic value between 0 and 10 (0 being “bad” and 10 being “good”). Kudless is concerned about the implications of this, saying: “It essentially prejudges the images that are being produced. Midjourney has a setting which forces the images to have a very painterly effect. Even if you put gibberish into Midjourney, it will still produce a beautiful image. You can search the database based on score. There are literally no images between 0 and 1 — there are no images that have gotten the lowest score possible, which is odd to me. And there are also no images any higher than, like, 7.5. So there’s this mythical, absolutely ugly image and this other mythical, absolutely phenomenal, beautiful image, and no known such images exist apparently, according to this algorithm…. That worries me a lot, because if you can’t produce something that’s ugly, there’s just this negative feedback loop we’re stuck in.”

The opposite approach is something like Stability AI’s Stable Diffusion image generator, which is completely open source. Born of the libertarian culture of Silicon Valley — the framework Wikipedia is built on — anyone and everyone can (and should!) access, edit, and retrain the model. 

Kudless explains: “That’s the only way to democratize access to the data. If the model is biased, then they would agree. But now you have the tools to retrain it and make it your own vision. Overall, I agree with that, and I’ve been using Stable Diffusion much more because of that accessibility. On the political side of it, I lean much more on the side of giving access of the technology to people, because the more people that are arguing over it, testing it, and pointing out the problems, I’d say the better. I’d rather have more people doing that than a closed group of a couple dozen from Microsoft or OpenAI deciding what goes in the model and what doesn’t. No matter what, it’s going to be problematic, but I’d rather deal with an open system where we can see what’s happening and argue over the problems than a proprietary system.”

Issues of ethics and data bias must be at the forefront of any conversation about AI today. As Kudless says, it’s going to be messy and problematic either way, as democracy always is. But when it is not accessible, it could be outright dangerous. Massive amounts of data in the hands of a corporate few, with the express incentive of generating the most profit possible, is a recipe for failure, including heightened economic inequality, racial injustice, and gatekept power and decision-making.

But AI is not, and may never be, a finished product; it will be what we make it to be, and it will be used for what we use it for. And though it is fraught with issues, we can’t disengage from it. Fleischmann says: “I don’t think we can dig our heads in the sand and try to ignore the potential hazards of AI that is designed, developed, and deployed in unethical and inequitable ways. Everyone in our society, from academics to journalists to lawmakers, has a responsibility to consider the potential implications of AI, and to thoughtfully consider how AI can be designed, developed, and deployed to benefit society.” 

Architects need not be programmers or computer scientists to engage with these technologies now and to be part of that societal consideration Fleischmann advocates for. Our essential expertise is to envision what things will be. A broader definition of architecture moving forward may include not just constructing physical buildings, but imagining the environments — physical, digital, social, political — that will steer us toward a better collective future. Those environments shouldn’t be, in the words of Saltz, “derivative and familiar” spaces that double down on the current modus operandi that has led to climate crisis and record levels of inequality; we could instead imagine what not just post-AI art, but a post-AI world would be when it “provides its own vision and vocabulary, its own sense of space, color, and form.” We could explore the optimistic, the unknown, and the post-work future AI might bring about. 

Davis Richardson is a licensed architect in the state of Texas and works at REX in New York City. He is a regular contributor at The Architect’s Newspaper and has taught at NJIT and the Architectural Association.

Leave a Comment