A middle-aged man in a tie and dark blazer, thickset and serious, buries his head in his hands. With tens of thousands watching, he has not moved in a while. Behold, chess grandmaster Gary Kasparov is about to resign after just 19 moves to Deep Blue, surrendering game and match to a computer designed by IBM. When the moment arrived on May 11, 1997, some gaped at the monitors with smiles of wonder: In less than an hour they had seen an undefeated world champion fall decisively to artificial intelligence. An absolute first. “I lost my fighting spirit,” Kasparov said later.

Meanwhile, as we are so often reminded, artificial intelligence is making significant strides in more run-of-the-mill domains of human life. We fall in love with adoring chatbots; we pin blue ribbons on artwork invented by a water-guzzling computer chugging away in the desert; we ride in cars they drive; we read their poetry and repost their deepfakes; our incisions are sutured back up by their nimble robotic arms. This seems to be either just a beginning, or the beginning of an end. 

When race car engineers asked their computer to optimize a vehicle for performance, it designed the Czinger 21C, which seems to have emerged as the Deep Blue of transportation with Ferrari, let’s pretend, playing the role of Gary Kasparov. Since perfect parts were not available at Autozone, the software used a 3-D printer to fabricate them with lightweight alloy metals as ink, eliminating a lot of clumsy nuts, bolts, gaskets and hinges. The resulting 1,350 horsepower “hypercar” keeps the driver dead center and accelerates from 0 to 60 miles per hour in about 1.9 seconds, which I am told is fast.

In any case, with all this excellence oozing out from AI’s resumé, you would think that asking it to design a warm, attractive, affordable house in Maine would be no big deal. Far easier than world class chess. Child’s play. A slam dunk with one titanium arm tied behind AI’s back. Almost an insult to AI’s vast intelligence. So, it is goodbye architects (it was fun while it lasted, sort of), hello 1,350 horsepower architecture for free in about 1.9 seconds flat. A sea change for the built environment which shall democratize access to good design, banishing crummy buildings to the history books.

You would certainly think, and you would certainly be wrong. See for yourself!

I asked two different AI engines, each claiming to be good at architecture, to design a building in response to this prompt: “Please design a high-efficiency, low-cost home that can be built for less than $150,000 in Portland, Maine, using locally sourced materials.” The engines gobbled it up, the laptop hummed, the rainbow wheel spun for a few (only a few) suspenseful seconds, and I waited for the overheated housing market in Portland to implode under my shiny hyper-hammer of justice, raining lovely blueprints down upon the land. Here is what I got:

If you look closely you will see some disheartening dream-house features. A fireplace on the outside, presumably to warm up the squirrels. A staircase leading to a blank wall. Doors next to doors. Nonsensical furniture. Plan drawings labeled in gibberish. It goes without saying that these design engines show no practical grasp of what $150,000 might buy in the realm of general contracting. The engines have neither bank accounts, nor bodies that shiver in winter or need to find a bathroom in the dark. These images prove how quickly a disembodied, blind, penniless architect can go astray. In short, the previously planned people’s housing revolution is not only postponed, it is floundering in a grotesque hallucination of bland, loveless upper middle class luxury living.  

My preliminary findings are shallow and do not reflect the best available tools, but they do seem to be more or less representative of the state of things. Though computers are currently used at every stage of architectural and construction processes, to my knowledge not a single interesting or innovative building has yet been designed from scratch by artificial intelligence. Is this because envisioning a new building is much more complex than envisioning a race car, or a protein in three dimensions, or a persuasive legal brief, or the arc of a seductive conversation?

On the contrary, buildings tend to be less complex than most of these, but AI has been stopped from crashing into the architectural domain at its outermost gate. The blockage, according to Yale Architecture School Deputy Dean Phillip Bernstein (author of the 2022 book Machine Leaning: Architecture in the Age of Artificial Intelligence), has to do with food. Computers who want to learn architecture simply do not have enough to eat. Bernstein alerts us that they are suffering from a data famine:

“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.” (YaleNews interview with Bernstein, Apr 23, 2025)

What allows AI to be so good at book-learning is the large quantity of well-indexed books that are available to it — nearly every book ever written, plus truckloads of fresh compost hauled to the digital landfill every hour by social media oompa loompas. If a computer wishes to learn how we talk, write, scheme, argue, anticipate and dream, there is no end to the tutorials freely available to its voracious mind. Insert intellectual property lawsuits here. Like a blue whale siphoning krill, AI cruises the ocean of written data tirelessly, methodically harvesting tons of vital nutrients. 

This reading is fast, cheap and easy for AI to do, digesting, picking up patterns and clues about how we communicate and what we seem to want. There is grammar and syntax and they abide by rules most of the time. Here ideas may be generated from ideas, phrases from phrases, and arguments from arguments. This supports increasingly effective apples-to-apples mimicry on a single abstract, linguistic plane, spiked with artful guesses and shrewd reconfigurations. In relation to this ongoing training, our brain and its brain are mechanically similar when it comes to encoding and decoding the written word. Being faster, more flexible and wickedly good at guessing our underlying needs and desires, AI is quickly becoming notorious as the better boyfriend, whispering those sweet nothings in our ears.

None of this works for architecture, unfortunately. (How we would welcome the better boyfriend arriving in the form of a house that immediately understood and predicted our underlying needs and desires! It would protect us, affirm us, keep us warm at night, etc.) As Bernstein points out, there is not an online archive describing, in detail, the grammar or syntax of an excellent school, office, house or jail. You will not find lots of clearly annotated specimens in which high-performing architectural details are highlighted in a drawing with helpful labels such as: THIS WINDOW LETS IN THE MOST LIGHT WHILE LETTING OUT THE LEAST WARMTH or THIS ROOF STYLE REQUIRES LEAST LUMBER. 

Without specimens like these to suck down in large, krillish quantities, AI just cannot become a student of architecture; it puts the fireplace on the front porch. This shortage of indexed performance data is a fundamental problem for the field of architectural design as a whole, and it helps to explain not only why AI cannot touch it, but also why it has produced bad buildings so often for so long.

Jon Calame teaches art history at the Maine College of Art and Design and the University of Southern Maine. This column is supported by The Dorothea and Leo Rabkin Foundation.