Article by Ridhi Purohit, and Judah Axelrod: “Before building a new house or apartment building, residents and developers must first ensure their project adheres to local zoning rules.
But zoning documents are often difficult for both applicants and local government staff to navigate. This can lead to extended back and forth between staff and applicants, slowing down the permit process at a time when most communities don’t have enough housing.
Some local leaders have expressed interest to Urban researchers in the promise of using generative AI to make zoning codes easier to understand, whether through screener tools that scan permitting applications to ensure they are complete, or chatbots that can answer development questions. However, leaders are hesitant to adopt tools that haven’t been properly vetted for the quality of the information they produce.
To test how well generative AI tools could interpret zoning codes, we ran a benchmarking exercise that evaluated the capabilities of various large language models (LLMs), building on previous work that explored whether machine learning could help automate the collection of standardized zoning data. To do this, we developed a set of zoning and permitting queries for Minneapolis—a city with a complex, 467-page zoning code and with zoning processes familiar to our team…(More)”.
