Artificial intelligence (AI) technology is automating tasks that once were the sole domain of human beings. AI-powered machines are diagnosing heart conditions, predicting the weather, and even writing screenplays.
Could computers ever design a safe, beautiful, and fully functioning building? Probably not, says Phillip Bernstein, deputy dean of the Yale School of Architecture.
Formerly a vice president at Autodesk, a major software provider for architecture and engineering, and a licensed architect, Bernstein calls himself a measured skeptic of AI’s potential, though he acknowledges that the rapidly advancing technology will change how architects perform their work.
His 2022 book, “Machine Learning: Architecture in the Age of Artificial Intelligence” (RIBA Publishing), lays out a strategy to help architects take advantage of AI innovation without being supplanted by it.
Bernstein, who is working on an updated edition of his book, recently spoke to Yale News about AI’s potential effects on architecture and the built environment. The conversation has been edited and condensed.
Phillip Bernstein
Before we discuss AI, how has previous technological innovation changed the way architects work?
Phillip Bernstein: Architects primarily use technology for representational purposes. We design and describe big, complicated things using images, drawings, and models. For thousands of years, we used a notational system on paper — that is, we drew stuff. Then from roughly the mid-1980s until the mid-2000s, we used computers essentially to do the same thing — plot stuff on paper. Using computers to make drawings took less time and was more precise than doing so by hand.
At the same time, buildings were becoming more complex, so we were drawing a lot more. A recent article in Harvard Design Magazine had a great infographic that showed how many drawings it had taken to represent buildings on Harvard’s campus. A building built in 1718 took two drawings. Another built in 1911 required about 20 drawings. By 2016, it took more than 1,000 drawings to depict a new lab building.
Around 2008, digital modeling became adopted through what’s called “building information modeling” (BIM), which enabled us to make behaviorally interesting three-dimensional models of buildings. BIM enhances precision and efficiency because it manages all the data about the building and uses that data to generate the drawings, but it doesn’t remove the architect’s agency. A human being is still making all the decisions.
How are architects currently engaging with AI?
Bernstein: Right now, the industry is largely using platform generators like ChatGPT or Claude to perform small tasks. For example, you can take your firm’s marketing material and train a language model on it so that you don’t have to write proposals. A lot of startups are doing all kinds of other things with AI, some of which is working, most of which is not.
What’s working?
Bernstein: The most interesting applications I see are the ones that make our jobs a bit more efficient, like those that document and analyze certain aspects of the work of translating a design into a building. For example, there’s AI-based software that takes a building information model and analyzes its compliance with building codes. It does an incomplete job of it, but it’ll get you started on the easy stuff.
There are platforms in development that can help translate design data into construction data. They take a BIM and convert it into what are called construction means and methods — the strategy for building a building: What’s the labor force? How many crews do I need? How much is this going to cost? How long is it going to take?
These developments are all very interesting, but I haven’t seen anything that could cause a fundamental disruption in the way architects work. It’s very hard to disrupt a profession that can trace its origins to the Enlightenment.
Bernstein and colleagues at the School of Architecture provided an identical prompt to four image generators, which produced four distinct buildings.
Prompt text: “A high-resolution perspective view of exterior modern apartment based on those of the 1970s with one level of living and many large glass windows. only one main level. ranch style, dark gray siding and mullions, large windows, think urban in a suburban area, trees in the background, mid-fall, make the entryway inviting and well-lit, fit the entire home in the view, crisp straight lines, add a couple people randomly walking inside the house, and one in the foreground enjoying the landscaping, the landscape grounds are meticulously maintained.”
What innovation might disrupt the profession?
Bernstein: In my view, what we would need is a ChatGPT for the building industry — a fundamental model that can reason about complex, three-dimensional objects that operate over time, like buildings. And that’s extremely hard to do because tools like ChatGPT can’t reason in three dimensions, or temporally.
Here at the School of Architecture, we did an experiment with Vincent Guerrero, who’s our senior director of advanced technology and a campus leader for AI on the IT side, in which we wrote a highly detailed prompt describing a building, and then gave it to four different image generators. The resulting images were very different from each other and only shared general characteristics like the choice of materials — concrete, glass, etc. So, tools like that can be helpful in terms of sparking ideas: you could ask for 10 pictures for a three-story, concrete-frame building, and it may show you things that you hadn’t thought about.
But, as I said earlier, these tools don’t understand what a building is. If I talk to an AI platform and say, “Okay, now make the floor-to-floor heights 10 feet, 7 inches, and I want spandrel depths at 18 inches instead of 14 inches, and make that glass slightly less reflective,” it might be able to complete the third task because it’s a manipulation of the pixels on the screen. But it can’t address the design of the building itself.
People talk a lot about multimodality in AI — tools that encompass language, images, voice, and video. A building is the most multimodal thing you can imagine. It has complicated spatial characteristics. It has materiality. There’s a whole logic around designing it, a different logic around building it, and a third logic around using it. This makes creating the built environment a very complex enterprise and we are nowhere near having an AI platform that can reason through all of that. We’re just playing around in a middle zone right now.
How would we move from the middle zone?
Bernstein: In my book, I argue that there is going to be a liminal period between BIM and useful AI that I call the “data interstice.” Right now, there are hundreds of pieces of software in the building industry spitting out data — we’re digitally generating all kinds of drawings, models, spreadsheets, and reports — but there is no coherent theory of knowledge for organizing it or making it useful training data for AI.
It’s easy to find a legal precedent if you’re an attorney because all the cases are indexed in databases. In the building industry, our data is incredibly disorganized and heterogeneous. We’d need a strategy for organizing and homogenizing it. From there, we might be able to build that AI model that can reason in three dimensions.
What do you think AI will ultimately mean for the profession?
Bernstein: I think there are aspects of the architectural enterprise that will be accelerated, augmented, and in some cases, possibly replaced by AI. But I do not believe that machines will ever be architects.
Architects have several responsibilities beyond simply designing buildings. First and foremost, we are responsible for public health and safety, which is why we are licensed to practice. We communicate with our clients to understand how they envision a project. Then we work with builders to translate our design ideas into reality. Can we really delegate these professional responsibilities to an algorithm? Would society want us to do that? If a project goes badly, a client can sue me because I’m professionally responsible for it. If we were to allocate this responsibility to an algorithm, then that algorithm better have a good insurance policy.
That said, if Sam Altman [the CEO of OpenAI and a leading proponent of AI’s potential benefits] is right — and a lot of other people are wrong, including me — and we eventually achieve artificial general intelligence that meets and surpasses human intellectual capabilities, then the entire world will fundamentally change, including this profession.
