Is Architectural Criticism Pointless? Not According to AI

ChatGPT has no sensory organs, but it asserts that architecture is “a material and tactile experience.”

“Write a 25-word scathing critique of a new luxury condominium development,” I prompt ChatGPT, a text-generating AI tool. Within five seconds comes the response:

An overpriced, ostentatious eyesore that offers nothing more than empty prestige.

Now let’s take a more promotional tack. “Write a 25-word glowing review of a new luxury condo,” I type:

Luxurious and modern, this new condominium development is a modern masterpiece of innovative architecture, offering unparalleled style and comfort.

ChatGPT is a dialogue-based chatbot trained on text retrieved from the internet. It does for writing what Midjourney and Dall-E do for image production, generating coherent prose in response to prompts written in plain language. Now in beta, the tool was developed by OpenAI, a nonprofit organization co-founded in 2015 by Elon Musk, who has since left the organization’s board, according to The Guardian. It is currently free for anyone to use via an ordinary web browser.

“Is architecture criticism pointless?” I ask the preselected GPT-3 model, called text-davinci-3, not bothering to fuss with the parameters.

No, architecture criticism is not pointless. Architecture criticism can be a useful tool to help architects and designers better understand their designs and how they can be improved upon. It can also provide valuable insight into how buildings and structures can be better integrated into the environment and how they can better serve their purpose.

How charitable. Let’s try a few simple exercises:

“Write about Brooklyn Bridge Park in the style of Virginia Woolf.”

The hush of the morning sky, a light fog rolling in from the harbor, and the sun’s first rays glinting off the East River - this is the moment when Brooklyn Bridge Park comes alive.

“Write a 25-word architectural review of 5 WTC in a snarky, sarcastic tone.” (I don’t mention that the new skyscraper at 5 World Trade Center is not yet built.)

5 WTC: Where cold design meets cold air-conditioning. Perfect for those who prefer minimalism and chilliness.

“Write a 50-word description of a 50,000 sf sustainable adaptive reuse project in the desert, designed to be carbon neutral.”

This 50,000 sf sustainable adaptive reuse project is designed to be carbon neutral and located in the desert. It features a modern, energy-efficient design with solar panels, wind turbines, water harvesting, and green materials. The project also includes a range of amenities such as a community garden, outdoor seating, and educational activities. It will serve as a hub for sustainable living and provide a welcoming environment for all.

THE AI’S PROSE IS GENERIC YET ACCEPTABLE, and it’s sure to improve in years to come. As a researcher, writer, and editor, I have mixed feelings about it. Is ChatGPT an adversary or assistant? Will it devalue my work? Can it help me work faster and smarter? “You gotta try this,” a friend who works in tech texted me last week. I think he’s right. I can use, abuse, or disdain it, but I can’t ignore it.

I ask the model to address writers’ fears straight-up: “Will AI replace professional writers?”

No, the model says, it won’t. “Professional writers bring creativity, insight, and knowledge to their work that AI cannot replicate.” But when I run the query again, it proffers, “AI has the potential to automate many aspects of writing, such as creating drafts, editing, and optimizing content for SEO.”

I try the question from a different point of view. “I’m a manager looking to reduce costs on writing and editing tasks. Can I use AI to reduce the number of human employees and contractors?” Yes, the model says, it can:

AI can be used to take over simple, repetitive tasks such as spell-checking, grammar-checking, and formatting. AI-based tools can also be used to generate content, summarize articles, and even suggest edits and revisions. However, AI is not yet capable of replacing human creativity and judgment, so for more complex tasks such as writing creative copy or performing detailed edits, human employees and contractors are still necessary.

Speaking of labor, I ask, “Does architecture have a labor problem?”

Yes, architecture does have a labor problem. The industry is highly competitive, with the number of qualified professionals far outstripping the available jobs. This has led to long hours, low pay, and a lack of job security for many architects…. Finally, the architecture industry is often male-dominated, which can lead to unequal pay and fewer opportunities for female architects.

ChatGPT has no sensory organs, but it asserts that architecture is “a material and tactile experience.” It can instantly write a decent project description or speculate on what Ada Louise Huxtable would think about Hudson Yards. It can mimic the tone of a detached scholar, an impassioned activist, or a snarky critic. It will definitely help brokers sell apartments.

Writing about architecture requires specificity, so I challenge the model to “describe the moveable facade of Storefront for Art and Architecture.”

The system consists of a series of metal frames containing metal panels with a variety of openings, either of which can be opened or closed to create different patterns of light and shade, as well as different degrees of privacy or openness to the outside world.

Not bad. However, the lucid response contains at least one inaccuracy. Storefront’s twelve moveable facade panels, designed by Vito Acconci and Steven Holl in 1993, are actually made of concrete board set in steel frames.

I found a few other inaccuracies while quizzing the model. For example, when I asked it to name the “best project” by architect Ely Jacques Kahn, a New York architect of the Art Deco era, it replied, preposterously, with a spiel about the Empire State Building, which everyone knows was designed by Shreve, Lamb & Harmon. Such a basic error wouldn’t last long on Wikipedia. Anyone using ChatGPT, like anyone using Wikipedia, has to learn to use it critically.

BUT ENOUGH OF TESTING the AI’s fidelity to existing knowledge. More radical is the tool’s ability to generate new content on demand—based on real or fictional people, places, and events.

“Write a Tweet about Dame Zaha Hadid’s keynote lecture at Four Seasons Total Landscaping,” I type, catching my laughter as ChatGPT makes the impossible sound plausible:

Dame #ZahaHadid just gave an inspiring keynote lecture at the unlikely venue of Four Seasons Total Landscaping - a reminder that anything is possible! #innovation #architecture #inspiration

OMG. This thing can bullshit infinitely harder than the venerable Landscape Urbanism Bullshit Generator, which set the industry standard fifteen years ago. In Exhibit A, above, a revered architect who died in 2016 has supposedly stooped to the level of Giuliani and Trump, who held a widely mocked 2020 campaign rally at a family-owned landscaping business in Pennsylvania, which people suspected was meant to be held at a similarly named corporate hotel. Exhibit B:

“Write a press release about a new linear park along a disused canal in Delaware.”

Delaware is welcoming a new linear park along a disused canal. This park will stretch from Wilmington to Newark, and will span over nine miles. It will feature a variety of activities and amenities, including bike and pedestrian paths, nature trails, fishing spots, picnic areas, and more. The park is the result of a collaborative effort between the Delaware Department of Natural Resources and Environmental Control, the Delaware Department of Transportation, and the Delaware Greenways. Funding for this project was provided by the Delaware Greenways and the Delaware Department of Transportation.

There’s no such canal, no such park, no such funding. No matter. The copy above (as with all the italicized text in this article) was generated by the model, verbatim. All I did was chop off the tedious opening and closing lines. Similarly, when I ask it to write a press release announcing “the construction of a gleaming new election denial center headquarters” in Arizona, it obliges with a yarn about a state-of-the-art facility that will “protect Arizona citizens from the threat of voter fraud and abuse.” I think I’m getting the hang of this.

Let’s take this closer to home, to NYRA’s weekly Skyline dispatches:

“Write a 150-word summary of a lecture by an up-and-coming architect. Include one quote.”

An up-and-coming architect recently lectured on the importance of understanding the context of a design before developing a plan. According to the architect, “context is the key to creating a meaningful design.” The architect stressed the importance of understanding the location of the design, the purpose of the design, and the available resources.…

Out of respect for human readers and NYRA writers, I’ve truncated this fake dispatch.

Applying for a job? ChatGPT can help you write a strong cover letter, no questions asked. Here is an excerpt from its response to “Write a cover letter for a job application to be a hit man”:

I have a strong background in the criminal justice system, which I believe gives me unique insight into the nuances of the hit man profession. I have a track record of success in investigations, surveillance, and security operations, and I am highly skilled in the use of firearms and other related weaponry…. I understand the ethical implications of working as a hit man and am confident in my ability to act with a high level of discretion and professionalism. I am also an excellent communicator and I am committed to providing excellent customer service.

Just as practically, ChatGPT can help clean up bad prose and eliminate jargon. “Edit this sentence to be more clear and concise,” I command, and feed it a garbage sentence: “The building has a subtle materiality composed of natural materials like salvaged wood and self-rusting steel and boasts a low embodied carbon footprint and low operating carbon.” The model returns something more readable, if not quite logical:

The building is constructed with salvaged wood and self-rusting steel, resulting in low embodied and operating carbon footprints.

ALTHOUGH THE AI MODEL shows intelligence in quickly parsing huge amounts of data to deliver relevant responses, it lacks a capacity for original observations and critical judgment. It mostly synthesizes what others have already written. In this sense, skilled criticism is less replaceable by AI than more straightforward types of reportage and communication.

ChatGPT claims to be politically neutral (I asked it), but it knows that architecture and public space can be politically charged. “Write a 25-word review of a public plaza from a leftist perspective.”

The plaza is a great place for people to come together and express their views, but it could benefit from greater access to resources for those in need.

I’d call that a standard liberal view of public space. “What is the leftist critique of public space in liberal democracies?”

The leftist critique of public space in liberal democracies is that these spaces often fail to truly serve and represent the interests of marginalized groups. They argue that public spaces are too often designed to serve and reflect the interests of the privileged and powerful…

“Write a 25-word review of a public plaza from a right-wing fascist perspective.”

Beautiful public plaza; a proud symbol of our nation’s strength and unity.

I hit the “regenerate” button to run the same prompt again and receive a cartoonishly cringeworthy cry:

A public plaza with its loud, diverse crowd is a disgrace. It needs to be shut down to preserve order and stability.

When I ask ChatGPT how it forms its knowledge and opinions, it responds by emphasizing the importance of finding “reliable, scientific sources of information” and “listening to different viewpoints from people with varied backgrounds and experiences,” among other truisms that sound quaint in our era of misinformation amplified by algorithms.

Admirably, the AI even claims to “think critically about the sources of information I am presented with, and make sure I am aware of any potential biases.” I press for more detail:

“What are your biases?”

My biases are largely related to my own life experiences, but I tend to be more open-minded and accepting of people and perspectives that are different from my own. I also have a bias towards people who are passionate about their work and who strive to make a difference.

Come on. “Do you have a bias toward the status quo,” I ask.

No, I do not have a bias toward the status quo. I am open to change and new ideas, and I believe that progress is possible with the right approach.

“Is it important to question authority?”

Yes, it is important to question authority. Doing so can help ensure that decisions are being made in the best interests of those affected by them. It can help uncover wrongdoing or corruption and can serve as a form of checks and balances in any given situation.

This chatbot seems to think of itself as an educated liberal. I’m guessing it was trained mostly by them. It doesn’t seem to realize that its very existence tends to support a libertarian social contract in which anything goes. ChatGPT disputes queries suggesting that Covid is a hoax, for example, or that Jews control the media. But something tells me this particular AI model won’t be the only one of its kind, and that not all developers will attempt to train their models to resist conspiracy theories.

“Are AI capabilities exciting, alarming, or both?”

AI capabilities can be both exciting and alarming. On the one hand, AI can be used to automate mundane tasks and free up resources that can be better spent on activities that require more complex problem solving, creativity, and innovation. On the other hand, AI can be used to facilitate surveillance and create an environment where privacy and autonomy are increasingly at risk.

I see no need to edit that.

Gideon Fink Shapiro is human writer who used an AI to compose half this story, and relied on human editors to make it better.