In the spirit of being a little vulnerable, one of the first things that I did in approaching this piece was typing my full detailed assignment into ChatGPT to see what sort of article it came back with. “Let’s really give it the best chance at doing this,” I thought as I refined my prompts, staring my own professional demise in the face. “Maybe I could be a massage therapist or something.” The several articles that I generated with AI are not what you’re reading now; and I’m not going to pull that trick of suddenly revealing that the past five paragraphs were written by ChatGPT. What you’re getting here is all me. We’ll talk about why in just a minute.
If you haven’t already heard (and you’ve probably heard by now), ChatGPT is an artificial intelligence (AI) chatbot developed by OpenAI. It’s in the process of upending the world, kicking off an AI development arms race, and redefining everything from how middle-schoolers complete their homework to how BuzzFeed authors its articles. Based on prompts entered by the user, ChatGPT can generate text in an instant. It can also answer complicated questions with relative directness and proficiency (though it still tends to occasionally get some facts wrong). To my unbridled joy, ChatGPT-3 can immediately write you a serviceable rap diss track about the time that “hedonistic sustainability” architect Bjarke Ingels met with former Brazilian president Jair Bolsonaro (though this type of thing seems against the more-mature-but-less-cool judgement of the more-recently released ChatGPT-4). These products are, without hyperbole, world-changing technologies.
Immersing myself in all this, one of my first instincts was wondering just how effectively ChatGPT could be used for architectural proposals, with their relatively formulaic writing and by-the-numbers content asked for by clients over and over. These documents are (as I have written before) painful to create. Beyond being entirely overhead, potential clients that you cannot control dictate your production schedule. It makes for a difficult job. Why wouldn’t you use robots to make your life easier?
Throughout early March of 2023, I spent time with ChatGPT-3 (model text-davinci-003), learning what it could do, and reading about different ways you could improve the prompts given to the AI. I listened to some instructional podcasts. I watched a bunch of “How To” YouTube videos and poured over OpenAI’s corporate content. I tried generating different types of content, coming at it from different angles, refining prompts to make outcomes into something more and more workable from the outset. I then revisited my experiments wholesale with the release of ChatGPT-4, the much improved “Mar 14 Version” that came out during the time this article was being written.
After some playing around with each version, I made an earnest attempt to start generating written proposal content. I imagined myself working for an amalgamation of my former employers, a fictional firm whose information and history I knew well within my own mind. I fired through the cookie-cutter elements in an architect’s written portfolio that are asked for all the time in RFPs: firm profiles, professional bios, cost control measures, sustainable design philosophies, project profiles. Making up some parameters for a fictional project that I could imagine fully (at least in the level of detail given within an early-days RFP), I also generated some attempts at writing for an approach and methodology.
Some of what the AI generated was fine. Without question, ChatGPT-4 is the better tool. With the right prompts, it is possible to have a reasonably well-done professional bio done in basically no time. I gave prompts for a fictional technologist whose resume I made up. With minimal effort (a few written prompts to clarify details), both ChatGPT-3 and ChatGPT-4 created writing that I’d have been comfortable using in a proposal or online. I think this would be a genuine time-saver for a busy proposal writer tasked with writing up a bunch of new staff profiles. You’re giving the AI the details it needs, and it sums them up in writing. Great start. Staff bios done.
The content that ChatGPT created for a “Sustainable Design Philosophy” was passible, but a little generic and lacking specificity. Same thing for cost control measures and for general descriptive profiles of an architecture firm. This content was also a stark reminder of how hollow the language of proposals and architectural marketing can sometimes be. Something that I noticed throughout this whole process is that ChatGPT’s answers tend to lean hard on things that have become cliché within architectural writing. Don’t we all love spaces that are “functional as well as beautiful”? Isn’t that the point of all this? I had a long think about how useful this content was. Would it set you apart as a firm? No, especially not if everybody in the world is also using ChatGPT. Do I think it would score well with evaluators? Not necessarily on its own, but with the right prompts, editing, and information to back it up, it could create something that could win. Have I personally submitted these kinds of clichéd answers when I didn’t have a better answer to give? Yes, but not happily. Sometimes the clichéd answer is also the true answer.
If nothing else, I think the content that ChatGPT generated shows the importance of firms developing real working procedures and policies in a way that distinguishes them from their competitors. It also shows that this type of refinement is a rarity, and not something that ChatGPT can magically make true for you, even if the AI knows what a pretty good answer is supposed to sound like. An example of this: on the first try with minimal prompting, both versions of ChatGPT wrote comprehensive responses when asked for a firm’s philosophy for meeting Canada’s goals for Truth and Reconciliation (something often asked for in RFPs for government-funded projects). The AI-generated answers were great, but the initiatives they described in the responses wouldn’t have been true of any firm that I had ever worked with. So, while the answers weren’t usable as a short-cut for submissions, they could easily be used as a template if you wanted to improve a firm’s actual policies.
Project profiles were slightly more difficult to get ChatGPT to do well. Even with well-thought-out prompts, the AI cannot see a photo of a finished building or walk through the space, so the writing always struck me as a little disconnected from the actual project (ChatGPT-4 was much better at this test, and the AI will eventually be updated to comprehend images—something not publicly available at the time of this article’s deadline, but it may be out by the time you read it.) The same difficulty is true for generating an approach and methodology, a common proposal requirement where specificity towards the project at-hand is usually your best friend. I can see how ChatGPT may be used as a starting point, but there’s a lot of work still to be done, no matter which version you’re using. ChatGPT can’t think about the potential client and what may win them over. ChatGPT may also volunteer services, design ideas, or meetings that may not be true for you—and while the AI can certainly be used to speed the pace of generating new content, it does not relieve you of the responsibility to edit and fact-check. If anything, it makes these steps much more critical. As it stands right now, the AI has disclaimers about generating incorrect or misleading content. It also clearly states that the AI is “not intended to give advice.”
None of the criticisms levelled above detract from the fact ChatGPT is an important tool, and there are times when it will prove invaluable. One of my biggest takeaways from my experimentation process is that the “rephrase” prompt is something uniquely powerful. It allows you to enter text and have it reworded in an instant. This is great if you’re not a fan of the first version of something that the AI created for you, if you need to shorten something, or if you want to just quickly experiment with the wording of something that you’ve already created. It’s also suddenly very easy to apply a different tone to writing that already exists. The unethical among us will probably use this feature to rip-off text wholesale from other designers, and make it instantly their own. Maybe I’m being judgemental on that point. If what the thief puts forward is uniquely generated and true of what they are proposing to do, then what difference does it make? Is it the same thing as being inspired by somebody? I honestly don’t know. Probably everybody reading this will have to reckon with that question in one way or another.
It’s also worth noting that proposals are far from the only application for ChatGPT in architectural practice. I spoke with Drew Adams, an architect and associate with LGA Architectural Partners in Toronto. He told me that he was experimenting with deploying ChatGPT to speed materials research, initially using the AI to try and quickly find suppliers in North America who may be able to make a specific mass-timber application, then also to find local material suppliers and rank them by price. Not all the results given back by ChatGPT were a home run for him, but using the AI had sped things along, turning the research into a process that was more like fact-checking.
I asked Adams if he had used ChatGPT to generate any actual written content for the firm: he said the answer was “no,” but that he had jokingly experimented with using it to generate a three-part specification for vapour barrier when talking with a colleague. “In thirty seconds, it spat out a three-part specification for vapour barrier. It seemed like it was probably as close to a rough starting block as you might download off a supplier’s website as a template.”
Adams’ descriptor of a “rough starting block” rung true for me in characterizing a lot of the content that I had generated, too. A question I kept asking myself throughout the process is: how much time and work would this actually save on a proposal? In my experience, a lot of proposals are generated off the strength of a firm’s existing marketing content, meaning you start with a lot of the writing you need, and it’s already tailored directly to your own business. The time spent on a proposal is mostly trying to edit and format things you have into something that speaks specifically and accurately to the project you’re chasing. I’m not convinced ChatGPT would have saved me all that much time in an actual proposal pursuit, assuming I was already equipped with pages and pages of firm-specific marketing material, though I’d have been happy to have it as an arrow in my quiver.
As for the many AI versions of this article that I generated in my nihilistic attempt to replace myself, they all suffered from many of the same issues I outlined above. Some inaccuracies. Some surface-levelness. They all delivered basic information but shied from depth, specificity, or explorations of how something felt (AI doesn’t know the feeling in the pit of your stomach when you’re working with less than an hour to submit something that could potentially be worth millions of dollars.) Having said that, I cannot deny that the article created by ChatGPT-4 was much better than the version I had created using ChatGPT-3, and that is probably my biggest takeaway from all of this. This is a technology that is improving rapidly, in an incredibly competitive environment, where some of the most powerful companies on earth have existential skin in the game. It will inform architecture firms as much as it will every other business—and it would be foolhardy to ignore how this may apply to your work.
Jake Nicholson is a writer based in London, Ontario, with extensive experience working on proposals for architectural and engineering firms.