What artificial intelligence (AI) thinks of historic preservation

Unless you’ve been sequestered over the past year, you’ve probably heard about the growing controversy on artificial intelligence (AI). A recent example is ChatGPT, a “chatbot” created by OpenAI. To use ChatGPT, you create an account and, through a web based “chat” interface, you ask ChatGPT questions using your natural language. You can, for instance, ask ChatGPT to “Summarize everything about the Apollo space program in 500 words,” and it will do this task for you, and reasonable accurately. More controversially, students can also ask ChatGPT to write an essay for them as part of their high school or college homework, and it does a surprisingly good job—”A” level work, according to some professors. How can this be? Have we finally entered the world of science fiction? Perhaps. But, more importantly, if ChatGPT, or any other AI program (or “bot”), can accurately represent the state of human knowledge, perception, and even values, what can we learn from AI bots about the world?

But first, let me provide some needed context for AI, and especially machine learning, for those who are not familiar with its concepts. Unlike a typical computer program, in which a human programmer has to predict all of the scenarios under which a program will perform, based on predetermined variables, an AI bot is programmed to access, process, and learn from data it is fed. Thus, an AI bot uses a kind of statistical process to predict outcomes; researchers feed data to the bot and the bot tests itself on predicting if certain variables lead to expected outcomes. If the bot correctly predicts an outcome, it keeps using the algorithm that it developed and improves on it, upon analyzing new data. In this sense AI bots use a positive reinforcement loop in that better predictions increasingly lead to even better predictions.

For example, lets say that you’re developing an AI bot that is supposed to “learn” from images, it is fed, what a “dog” is. You feed the AI bot millions of images of dogs and other animate and inanimate objects and reward algorithms that the AI bot develops that correctly identify a dog when the researcher provides dog images, but, conversely, also reward an algorithm that correctly fails to recognize dogs in images where one does not exist, thereby avoiding false positives. If you take this process to its logical extreme, you can develop an AI bot that will correctly identify the people, animals, plants, and things in an image, but more amazingly, you can ask the AI bot to create a new image that represents the visual characteristic it has learned to identify. Thus, the AI bot has become part artist, but more importantly, it has also become a window into how human beings see the world. And, what’s even more interesting is that we have many tools that do this kind of work today, including Midjourney, which is probably one of the most well known.

When you ask Midjourney to create an image of a dog for you (figure 1), it will create the prototypical image of what human beings think a dog is, but you can also ask much more of this AI bot. You can ask for it to create a “happy” dog being held by an “older woman” who has a “bemused” look on her face. What the AI bot will deliver for you, in terms of an image, will accurately represent how humans think a “happy” dog looks like; how humans think an “older woman” looks like, and what the emotion of “bemusement” looks like on this woman’s face. What’s quite eerie about this image is that the woman and the dog in the photo don’t actually exist; instead they represent a kind of schema in the algorithms that Midjourney uses to visually represent the world. This AI bot now knows more than any, single human being does about how most people—in vast aggregate—perceive, understand, and visualize ideas, concepts, and words. Midjourney is, in essence, a really important research tool that can tell you a lot about our world, today.

Now, let’s turn the power of Midjourney’s understanding of the visual world of people onto the field of historic preservation. If I ask Midjourney to create images for me about various aspects of the field, it should accurately and visually represent the concepts I ask it to render. Or, in a more crude sense, what does Midjourney think about historic preservation? Let’s find out.

If I simply ask Midjourney to create an image that represents historic preservation (figure 2), it consistently produces images of high style, Victorian-era mansions. No matter how many times I ask it to render, “historic preservation,” these are the images it generates. Note what is missing from these photos: Buildings that are not built in the Victorian era, commercial buildings, vernacular buildings, cultural landscapes, Indigenous landscapes, or more rural areas. Takeaway: For most people (in the US), historic preservation is primarily about residential, private property that consists of high style, Victorian-era mansions.

Now, I’m getting really curious. What if I ask Midjourney to put some people into a photo? I decide to ask the AI bot to create an image of “an employee from the state historic preservation office positioned in front of their work” (figure 3). Again, no matter how many times I ask Midjourney to render this image, it always contains White men, often in what look like uniforms (i.e., representing authority). Takeaway: Most people (in the US) think that employees who work at state historic preservation offices represent authority and are White and male.

One of the most common job titles in the historic preservation field is that of an “architectural historian,” especially within the regulatory sector of the field, which represents about three-quarters of all paid employment in the field. I ask Midjourney to show me an image of an “architectural historian” (figure 4). While there is clearly an antiquarian quality to the people’s dress, the result is rather similar to the SHPO employee, but the men skew older (and possibly more academic) in appearance. But, note how the buildings switch from residential to commercial or governmental. Perhaps this represents a sense of increased professionalism in the work of these individuals? Curious.

I have to admit that at this point in my exploration, I’m finding this AI bot rendering rather fun, but also sobering because it seems to be reinforcing the demographics of the field and its overemphasis on high style, pre-World War I buildings.

As a former academic, I can’t resist asking Midjourney to show me what it thinks an “historic preservation professor” is (figure 5). The results are rather disconcertingly close to my own visage — White, older, male and perhaps showing an condescendingly erudite pose. Yes, these people have something seriously important to tell you, no? (The gentleman at the lower left looks like he’s about to scold me for not paying attention or condemning me for my naivety on the proper description of the Classical orders.)

And, of course, the object of attention of professors is students. So, I ask Midjourney to show me what an “historic preservation student” looks like (figure 6). My reaction is literally, “wow!” This render could not be, from my perspective and from the data from the US Department of Education, more accurate: A young, White female. Yes, this is the image of the historic preservation student. Midjourney seems to be understanding the demographics of the field amazingly well.

Moving from people to agencies, this exercise begs for me to ask how Midjourney visually conceptualizes the National Park Service’s historic preservation work (figure 7). The results are rather surreal, but not so unfamiliar. That White man in the lower right? How bizarre, but all of the NPS’s visual archetypes are at work here. These are clearly images of greatness and power—from both cultural and natural perspectives.

And, of course, the other big federal agency in the historic preservation world is the Advisory Council on Historic Preservation. What does Midjourney think of the ACHP (figure 8)? Finally—other than students—it’s nice to see some women represented in these images, but again, we see lots of older, White men along with images that convey importance and power. And, the architectural styles are still in a high style, pre-World Word I mode. This makes me wonder how many of the images that the ACHP use represent more recent architectural history or vernacular landscapes? What about Indigenous landscapes? Where are those?

Lastly, I’m now really curious to ask Midjourney to show me a visual representation of the Secretary of the Interior’s Standards for Rehabilitation (figure 9). Again, nice to finally see a woman represented in the image, but as with many of these images, the visual concepts of power and authority are simply overwhelming and almost frightening. But, to be clear, this is how the public perceives the Standards and how federal agencies represent themselves (!), as conveyed by Midjourney.

Overall, the takeaway from this visual exercise is to recognize that the historic preservation field—as represented in its majority of paid, governmentally-driven practice—is what these images represent: White, male, older, highly educated and focused on power and pre-World War I, high-style buildings. We can say as much as we want to ignore this reality, but the point that is so crucial to understand from what Midjourney is telling us is that this is both what the public believes and the messages local, state, and federal governmental agencies are sending the public. On a more positive note, Midjourney can show us how our work, in the field of historic preservation, can and should change to be more relevant to more people.

Now I arrive at where I began, with the title of this article. My last figure is what I received when I asked Midjourney to render John Ruskin on the moon, in all its steampunk glory. Enjoy.