Architects, don't fear the AI
Machine Learning isn't coming for architectural design anytime soon
There is little reason for architects to fear that machine learning algorithms will replace them in the near-to-medium future. Humans will probably be designing buildings at some level for a long time to come. Here’s three reasons why:
Design is cheap to get right, and expensive to get wrong.
Architectural fees typically account for around 10% of a new building. More if it's a special or expensive building, less if it's a normal, off-the-rack design. Only about 10% of that 10%, or 1% total, is the "design" itself—the floor plans, the way it looks. The rest is coordination, documentation and quality control.
So, in terms of costs to developers, paying the architects the 1% of the project cost for the design is not something that will usually make or break a pro forma. Particularly because buildings are very expensive. Bad plans make construction even more expensive than normal and make buildings less attractive to tenants/buyers. They cause problems with zoning committees and banks. Any one of these can ruin a developer.
please don’t try to build this (satirical plan by Paul Keskeys)
Cheap to get right and expensive to get wrong is not a great problem space for machine learning. Rather, what has made these algorithms so useful is that in general they have a much higher quality/cost ratio than humans, and sometimes that ratio is more important than the absolute level of quality.
Think of the algorithm that determines what you see on your Facebook timeline. Suppose, instead of that algorithm, a close friend of yours spent all day searching the internet for stuff you might enjoy seeing, and sending you personalized links to the best stuff. Is your current Facebook feed as good as what that friend would create for you?
Obviously not. Your friend knows you much better than the algorithm does. But it’s not implausible that the algorithm might do the job 20% as well as your close friend. And at the same time, starting salaries for “your close friend,” who needs to eat food and live somewhere, are high compared to what Facebook needs to pay its algorithm, which only eats electrons and lives on a tiny hard drive. Electrons are cheaper than tacos, hard drives are cheaper than houses. So, it makes sense for Facebook to accept the large 80% dropoff in quality in exchange for the enormous 99.99+% dropoff in cost.
Note that this only works because “what you see in your Facebook timeline” is pretty low-stakes. Over time, in aggregate, the quality of what you see matters a lot for Facebook, but any particular mistake doesn’t matter that much. Did Facebook show you your great-aunt’s puppy pictures today instead of your old college roommate’s toddler pictures? Not that big a deal.
By contrast, if your computer accidentally specifies high-molybdenum-content marine-grade stainless cladding at $10/sqft instead of galvanized steel at $1/sqft for your client’s office building in downtown Indianapolis, your client will never speak to you again except in court. Protesting that your own time is billed at $100/hr+ while the computer made the decision for free will not help you.
Ok, you might, say, but what if machine learning algorithms aren’t just cheaper but better than humans, too? After all, not all machine learning solutions offer us the somewhat-worse/much-cheaper tradeoff. In some cases algorithms trained by machine learning have surpassed human skill. And, given the pace of innovation in the machine learning world, the number of such cases is likely to grow.
But it’s unlikely that designing buildings will be among them anytime soon, which brings us to…
Design success is hard to identify and always partial
It’s difficult to specify what makes a building design good, because good designs balance many different factors. Do all the rooms, or at least the important ones, have good light—not too little, not too much? Given the site’s location, do the windows look out onto pleasant views? Are the rooms well proportioned? What’s the ratio of unleasable to leasable square footage?
Buildings can be successes or failures on several of these fronts, but a good design is a synthesis of solutions to all these factors. Perhaps that’s just a way of saying “we don’t know enough about the design process to explain to a computer how to optimize it,” and that’s probably right. In itself, this would not make design a bad fit for machine learning. Distinguishing cats from other objects in the environment, for instance, is like this. You just have to have a lot of examples of cats vs non-cats to give the computer.
But success in distinguishing cats is different from design success in that it is easy to identify. You can give a computer 100,000 examples of cat photos and 100,000 examples of non-cat photos but try finding 100,000 examples of successful building designs vs 100,000 unsuccessful examples. Are any buildings just simple successes in the way that a picture is either of a cat or not? If not, that will make it very hard to train a computer to design successful buildings.
Architecture is a communicative medium.
Lastly, and perhaps most importantly, architecture is, in part, a way that people communicate with each other. A building's design can tell people: we are modern, we are sophisticated, we are expensive; or, it might tell them, we're old-fashioned; or, we're none-too-fussy. Algorithms will change what some of the terms in those messages are, but humans being humans, we will keep the channel open.
So, while it’s easy to look at the remarkable advances machine learning algorithms have made in the last few years, they’re not going to replace building design anytime soon.
Of course, building design isn’t the only thing architects do, and machine learning probably isn’t the last wave of “AI” that will frighten architects into thinking they’ll lose their jobs. Machine learning probably will come to architecture in some form (in small ways, it already is). Computers will continue to make architects more efficient, which may indeed threaten some jobs. And some as-yet unheralded form of AI may disrupt architectural design in a more profound way. But there is no near-term future in which machine-learning algorithms are designing buildings.
-David Schaengold, October 2020