One somewhat unexpected talent of ChatGPT and other artificial intelligence (AI) text generation tools is that they seem to be quite good at software development (i.e., writing code). Just as software provides instructions for computer behavior, zoning regulations provide instructions for permissible changes to the form and use of the built environment. However, while many planners refer to zoning regulations as "code," the syntax of zoning is closer to conversational English than Python or C++. And it's debatable whether subjectivity in zoning is a feature or a bug.
Putting aside the question of "should we automate zoning," are the current versions of AI tools capable of taking some of the drudgery of drafting zoning regulations? In the October issue of Zoning Practice, "Using Generative AI to Draft Zoning Codes," Norman Wright, AICP, attempts to answer this question through experiments with ChatGPT and Midjourney.
Zoning Text Mad Libs
As Wright points out, many planners already take a "Mad Libs" approach to drafting zoning text. When adding a section, we may look at the preexisting structure of the regulations and formulate new provisions that match that structure, or we may look to our neighbor's regulations for inspiration. In either case, our first draft focuses on the syntax that will help us address new policy objectives. The specific standards are, ideally, defined or refined through a participatory planning process.
Through experimentation with ChatGPT, Wright demonstrates that, with sufficiently prescriptive prompting, large language model AI tools are already capable of reproducing the syntax of a zoning ordinance. And that makes sense since imitation is their specialty. Automating this part of the drafting process may allow planners to devote more time and attention to the data collection, analyses, and deliberative processes that will help local officials identify the right specific standards to advance policy objectives.
Figure 1. Midjourney's response to the prompt, "Simple black and white isometric diagram of a 40 foot tall, 30 foot wide, 60 foot long rectangle on a 60 foot x 100 foot rectangle of land drawn with simple straight lines."
Wait a second, though. While good syntax is the foundation of legally defensible zoning regulations, contemporary zoning ordinances often lean heavily on visual aids to supplement or complement the zoning text. For example, it's often much easier for users to interpret the meaning of dimensional standards from a simple diagram than from a block of text or even a well-formatted table. And if a fairly prescriptive built form is central to the objectives of the zoning regulations, the ordinance will likely need dozens, if not hundreds, of precise illustrations.
This is perhaps the biggest area of opportunity for generative AI. For many planners, clear and compelling visuals are harder to create than formulaic zoning text. Wouldn't it be great to be able to prompt an AI image generator with zoning text and receive an elegant and accurate illustration in seconds? Unfortunately, though, Wright's experimentation with Midjourney ran into some technical difficulties. His results ranged from obtuse (Figure 1) to fantastical (Figure 2). With the rapid evolution of these tools, I suspect this is just a momentary delay and not a dead end.
Figure 2. Midjourney's response to the prompt, "Simple black and white isometric diagram of a 1,200 sq ft house on a 60x100 rectangle of land, dimensions included, drawn with simple straight lines."
Subscribe to Zoning Practice
Each issue of Zoning Practice provides practical guidance for planners and land-use attorneys engaged in drafting or administering local land-use and development regulations. An annual subscription to ZP includes access to the complete archive of previous issues.
Top image: An example of Bing Image Generator's response to the prompt, "A photo depicting zoning regulations" on October 10, 2023.
About the Author
David Morley, AICP, is a research program and QA manager with APA and editor of Zoning Practice.