This article is part of the Under the Lens series
Max Griswold is putting together a database of detailed eviction data for cities and census tracts in California. Last year, after making a batch of public records requests and outreach to sheriffs’ offices across the state, he received a blizzard of eviction records. But the documents were in rough shape; many had sections redacted, and they were in different formats, which made it difficult to compare them to one another. There was also the sheer volume of them—Griswold was looking at over 6,000 pages of court records.
For housing researchers like Griswold, this is a common problem.
“I can’t really feasibly read through that amount of court records,” says Griswold, who is an associate policy researcher at the RAND Corporation, a global, nonprofit research organization.
But in the past couple of years, a growing suite of AI tools has allowed researchers like Griswold to dramatically ramp up their efforts. “I was able to use vision algorithms to convert [those 6,000 pages] into actual usable data. And it took something that would have taken our research team two years of manual labor … into something that took a matter of days. It had a huge benefit for us in terms of trying to scale up that kind of work.”
What’s a Vision Algorithm?
A set of rules that train a computer system how to “see” and extract relevant information within a document.
Griswold’s project is representative of a patchwork effort of researchers across the country who are racing to put together databases that will allow them to finally get a clear look at what’s truly working (and not working) in U.S. housing policy. The more information they can gather, the better the chances of finding solutions to the housing crisis, they say.
AI tools like the vision algorithms Griswold used are helping analysts supercharge their research, and in many quarters, the pace of progress is increasing rapidly.
But the gap between AI hype and AI reality is sizable, and it still takes a huge amount of human labor to drive these projects. Even so, most researchers Shelterforce spoke with for this story are optimistic about progress, if only because there’s so much room for improvement.
Getting Good Data
It’s almost remarkable how much housing data we don’t have. There isn’t a centralized, national database that tracks housing in the U.S.—how much housing there is, the different kinds of housing available, who owns what, how much landlords charge for rents, how many evictions are taking place, and who’s moving where. We don’t even have a database of all the various (and voluminous) county and municipal zoning regulations that dictate what can and can’t be built, and where. Some makeshift local, statewide, or regional resources exist, but there’s no comprehensive national database.
One reason for this is that housing is an intensely local undertaking. Housing policy happens at the municipal or state level. There isn’t a lot of federal zoning law simply because what makes sense for single-family homes in, say, Long Island, wouldn’t really apply to lots in Nebraska.
“What are the actual barriers to building more housing supply? Who is bearing the brunt of the rental affordability crisis? What are the types of regulations that are getting in the way of building more supply? So many questions, I think, that people are trying to answer, and they have very simple answers right now because they don’t have good data,” says researcher Karen Chapple, director of the School of Cities at the University of Toronto, where she also serves as professor of geography and planning.
Jason Ward and Luke Schlake at the RAND Corporation are tackling one of the most basic questions of the housing affordability crisis: Why does housing cost so much more in one place than another?
For their analysis, described in a paper that was released in April, Ward and Schlake acquired reams of cost data for more than 100 multifamily housing projects in California, Colorado, and Texas—both publicly subsidized affordable, and privately funded market-rate housing—and analyzed what drove price differences.
For the publicly funded projects, information came via public information requests with state agencies that oversee funding for the Low-Income Housing Tax Credit program. To receive these credits, builders must submit detailed data.
For the market-rate housing, they lucked into a contact who was able to put them in touch with management at Trammel Crow Residential, a large for-profit builder. This was a big deal for them. “When you get into market-rate housing production costs, I think a lot of people consider those data to be proprietary in nature. So they don’t want to share them with anyone,” says Ward.
Right off the bat, they found that the data they received varied wildly in format and useability. “The Colorado data was beautiful,” Ward recalls. “It was just spreadsheets that were all filled out . . . And that was really easy. From Texas, we got a bunch of PDF data that was at least largely consistent in terms of having the same kind of format.”
But then there was the data from California, which was sent from various auditing firms. This data included different terms for the same line items, and the table formats varied.
“One of the big initial nuts we had to crack was how to translate all these PDFs of cost data into something that we didn’t have to literally just sit and visually copy over into a spreadsheet. Because that would have taken probably hundreds of hours. So there we had to do a lot of AI-assisted text recognition from nonstandard formats,” says Ward.
He used an Amazon Web Services tool called Textract to consolidate the data that was in the PDFs, and a RAND in-house large language model (LLM) tool to summarize some of the dense project requirements into more digestible bullet points.
What’s a Large Language Model (LLM)?
AI systems that use advanced statistics to uncover patterns in vast troves of written texts that they can apply to generate responses. [AP Stylebook]
But while these tools saved Ward and Schlake potentially hundreds of hours of painstaking labor, they are error prone. Ward spent a lot of time manually inspecting large amounts of the recovered text and making corrections.
Depending on the quality of the PDF, the font used, and other factors, a simple address like 122 S. Oak Street can be rendered as “1225 Oak Street” or “122 Soak Street.” Numbers can be a particular challenge. “You scan these data in and maybe it turned a period into a comma and suddenly now $1,040 is $10,400 or $104,000. So [a lot of the work] was checking to make sure all these numbers translated well, they all add up and make sense.”
This was a common refrain among the housing researchers we spoke with who use AI tools. Several stressed that you still need humans in the loop to closely oversee these projects. “If you don’t know what you’re doing, small errors can slip in, and it’s horrible,” says Quinn Underriner, a senior data scientist at the Terner Center for Housing Innovation.
Ward and Schlake’s project uncovered some intriguing and potentially actionable causes for skyrocketing housing costs. For example, they found that production timelines, which are associated with higher costs, are almost two years longer in California than in Texas, and that municipal impact and development fees added an average of $1,000 per unit in Texas, but an average of $29,000 per unit in California. While market-rate and affordable housing costs differed hugely in California and Texas, both types of housing cost about the same in Colorado.
They also found that a few pain points such as “unusually large architectural and engineering fees, likely related to highly prescriptive design requirements” were major drivers of housing production costs in California. In light of their findings, they made specific policy recommendations to alleviate some of the upward pressure on housing costs: a requirement for local jurisdictions to approve or deny development proposals within 30 days, synchronized building inspections to speed up construction timelines, and easing strict energy efficiency or design requirements to lower disincentives against building new housing, among many others. Whether these recommendations are taken up by local governments is an open question, but now legislators at least have access to potential solutions that, crucially, are backed by data.
The AI tools Ward and Schlake used may have saved them some time, but the work could’ve been completed without them. It helped, of course, that their project was relatively small, looking at only 144 multifamily projects across three states. (Ward and Schlake are already in the process of expanding to 11 states, with hopes to go national.) A project of truly national scope would require a lot more time, resources, and an exponentially larger role for AI. But is AI up to the task?
Can AI Do the Grunt Work?
AI tools aren’t just useful when it comes to analyzing the labyrinthine municipal regulations that govern what can and can’t be built in America—in many ways, they’re a necessity, say researchers.
Any researcher who wants a big-picture, national snapshot of the regulatory superstructure of our housing policy must ingest tens or hundreds of thousands of individual municipal codes, each of which can run to hundreds or thousands of pages. The New York City municipal code, for example, is approximately 12.2 million words long, equivalent to about 48,000 pages. Even given an infinite amount of time and resources, that’s just not feasible.
Of course, before you can hope to digest and analyze all those regulations, you need to gather them all into a centralized database, a task that’s nearly as daunting as reading them would be. Many researchers hope that AI can essentially take over this grunt work, independently scraping data from local government websites and compiling it into a database. This type of AI tool is sometimes called “agentic AI”—i.e., an AI tool that can exhibit some limited agency, meaning it can make some decisions and take some actions without being told to. This type of AI has been a holy grail of sorts for data researchers.
The good news is that it’s here; the bad news is that it’s not very good.
Existing agentic AI must be monitored and “handheld” to avoid errors, and it has trouble following basic instructions. The researchers we spoke with are reluctant to even use it.
“If the question is, is there some agentic AI model that you could send out and tell it, please merge all these data sets and create something for me? I would not use what came out of that,” says Underriner. “We’re not at the point yet where you could just be like ‘scrape all the HUD data, set it in a nice, neat thing.’ Because there are so many geographies and time periods and various program levels. And is it well documented? When was this last updated?”
[RELATED ARTICLE: Tenant Screening: A Billion-Dollar Industry with Little Oversight. What’s Being Done to Protect Renters?]
RAND’s Griswold echoed this assessment of today’s agentic AI. “Even just trying to get municipal ordinance language is a huge issue,” says Griswold. “Every single city is a little bit different. And so if you try to use an AI tool to help you to do that more interactively, the amount of nuance [needed] makes those tools currently not as useful for that purpose. Once you have that data in hand, AI is pretty useful to help you summarize it and get something meaningful out of it. But getting to that point is still a huge manual labor.”
Some researchers are making progress treating agentic AI’s work as just that—manual labor. In American’s Assembly Line, a book on the history of American factory work, author David E. Nye discusses early attempts at industrial automation. “The basic idea resembled the system used to produce piano rolls,” Nye writes. “As a machinist cut a part to precise specifications, a punched tape recorded his every movement. Afterward, just as a player piano can endlessly reproduce a sequence of notes, the machinist’s movements could be ‘played back’ as often as desired.”
At the Urban Displacement Project, where Karen Chapple is working on a database of anti-displacement policy in California, researchers are using a process similar to primitive “player piano” style automation to get their AI tools to work effectively. Chapple’s previous project involved gathering ADU (accessory dwelling unit) regulations in Canada which, like the U.S., doesn’t have a comprehensive database of municipal ordinances. During that project, Chapple soon butted up against the limitations of the AI tools she was using.
“We built a tool, and tested it, and found that if we asked really simple [yes or no] questions like, do you have a parking requirement with your accessory dwelling unit regulation . . . then it could answer it quite accurately and it could pull the data from many different cities,” Chapple says. “But if you ask a more complex one, like how big is the rear setback for mother-in-law cottages . . . It would get it wrong more often than not.”
Now she says she’s building on the work of Alexander Bartik, Arpit Gupta, and Daniel Milo, who recently came up with a novel method of using LLMs to compile large amounts of municipal code documents. According to their March 2025 paper, the trio downloaded the municipal codes of over 19,400 municipalities and over 16,000 townships, “vectorized” the documents (basically, broke them up and grouped similar sections together) with an OpenAI algorithm, fed them—in small digestible chunks—into a modified LLM, primed it with prompts like “you are a municipal zoning expert,” and then asked it questions like “does this zoning bylaw include mandates or incentives for development of affordable units?”
The consumer-behavior data people don’t know much about unbanked or unhoused people, because nobody wants to sell things to them other than payday loans.”
By instructing the LLM to “think out loud,” and provide citations for its answers, they were able to achieve high accuracy rates. One way that Chapple has improved on their 2023 results is by intentionally training AI on the work her undergraduate students completed manually when they created a database of 150 U.S. cities.
“We explain step by step what our students have done. And then you sort of look at the processes they use. And then from that you can extract something that’s repeatable—commands that you can give the AI tools.” Essentially, it’s the cognitive equivalent of factory automation: recording the actions of human workers to use as instructions to the machine.
But even explicitly modeling AI processes on human ones doesn’t produce perfect results. “The lack of accuracy in the tools is a problem,” Chapple says. “Now we’re getting 92 percent, 95 percent accuracy, but you kind of want to have 100 percent accuracy in your data.”
Chapple thinks the technology will improve exponentially in two to three years, at which point research will really take off. “If we could make the chatbot do the work for us of building these databases, it’s going to be pretty amazing,” she says.
But we’re not there yet—and we might not be as close as we think.
Rayid Ghani, a professor of public health at Carnegie Mellon, has direct experience with the limitations of existing data. Ghani believes that researchers who are pinning their hopes on [agentic AI] might be putting the cart before the horse.
“The bigger problem is going to be how you get access,” he says. “AI can help if [data] exist somewhere in an electronic way that are accessible on the internet. But if they’re not, if they’re hard to find, or they’re on paper, then AI isn’t really going to be able to do much there.” Using AI to scrape data is only feasible if data is scrapable, he says.
In the course of Ghani’s work, one of the biggest obstacles he’s faced is that there’s little data gathered on the populations he studies. One reason for this is that, in the absence of government action, a lot of data harvesting has been taken up by private companies who are motivated by profit.
“If you want data on [who] the marketing world cares about, you can get that data because they do spend a lot of time and money collecting it,” Ghani says. “Who lives where, who spends on what. But that’s not comprehensive. It misses out on a lot of people that we care about knowing. The consumer behavior data people don’t know much about unbanked or unhoused people, because nobody wants to sell things to them other than payday loans type of places.”
Early returns on the analysis side are similarly uncertain. While Bartik, Gupta, and Milo’s wide-ranging project is impressive on a technical level, one of their main takeaways is that “suburban regulations maintain exclusivity through density restrictions,” which is a little like saying you’ve used AI to discover that water is wet. As researchers use AI to compile larger and more comprehensive databases, there will no doubt be major findings and breakthroughs. But today, most of AI’s potential is still just that—potential, as yet unrealized.
Even if AI reaches its full potential, its power will be confined to the virtual realm of information. Researchers say any insight or analysis provided by an AI-assembled housing database will still require real-world support to be translated into real-world change.
“Obviously I believe in the power of research to ensure that programs are operating well and [getting] the outcomes that they are looking for,” says Underriner of the Terner Center. “But to be quite honest, even as someone who’s a researcher in this space, the bigger issue is resources and political will.”
About the Author
Comments