AI Seeks Out Racist Language in Property Deeds for Termination

California counties are under orders to find and redact language from historic property deeds that barred non-White people from owning homes, a process that can take years and cost millions.

But Stanford University researchers on Thursday unveiled an AI model they say can analyze decades of property records in just a few days at little expense, and they will offer the tool for free across the state and around the country.

Racially restrictive language in housing documents is outlawed and can’t be enforced, but it still exists in property records across the US. Stanford’s RegLab says that—working with Santa Clara County, home to Silicon Valley—it has developed a version of the large language model Mistral to detect these records. Once detected, they can be officially redacted.

One such covenant in a property deed says: “No persons not of the Caucasian Race shall be allowed to occupy, except as servants of residents, said real property or any thereof.”

“Our best estimate right now is that one in four properties in the county were subject to a racial covenant as of the 1950 housing stock,” said Daniel Ho, a Stanford law professor and director of RegLab, who led the project, which included collaboration from Princeton University.

A 2021 California law requires each of the state’s 58 counties to identify and redact racial covenants. But the costs and effort required to do so have so far been prohibitive. Los Angeles County, which recently began scouring its property records, said the process would take seven years and cost about $8 million.

Santa Clara County alone has 24 million property records, but the study team focused mostly on 5.2 million records from the period 1902 to 1980. The artificial intelligence model completed its review of those records in six days for $258, according to the Stanford study. A manual review would have taken five years at a cost of more than $1.4 million, the study estimated.

“By combining accuracy, speed, and cost-effectiveness, our solution offers a practical path for Santa Clara County—and other jurisdictions—to meet legislative requirements while preserving important historical records for further study and analysis,” the paper said.

Portions of a racially restrictive 1940 real property deed from Santa Clara County includes the language: “No persons not of the Caucasian Race shall be allowed to occupy, except as servants of residents, said real property or any thereof.”

Source: Stanford University

The researchers found that a handful of real estate developers were responsible for close to one-third of the restrictive covenants in Santa Clara County. Others refused to incorporate the racist language in the deeds, among them Joseph Eichler, a developer who built thousands of homes in Palo Alto and other parts of the Bay Area. “Eichler Homes” are part of the area’s lore, with Apple Inc. co-founder Steve Jobs among fans of the minimalist and simple design of the homes.

“This shows that small number of actors can meaningfully influence integration and fair housing,” Ho said.

Looking to the Future

Washington, Minnesota, and Texas are among the other states that have passed laws to let owners remove offensive or discriminatory language from historic home ownership records.

But such laws usually place the burden of removal on the homeowners, the paper said. “Between 1999-2021, California, for instance, maintained a process by which homeowners could petition with the County Recorder to modify a racial covenant on their property,” according to the study.

The model developed by Ho and his colleagues identified more than 7,000 records in Santa Clara County that had racially restrictive language. Some of these restrictive deeds can cover multiple properties, the authors of the study said.

“Our project demonstrates a path to using technology to aid in many other similar reform efforts that either never get done or take years to complete,” Ho said. “Consider another example: the number of congressionally mandated reports for agencies has ballooned over time, but it’s difficult even to know how many such reports exist in the U.S. Code, let alone determine which ones are ripe for sunsetting.”

AI large language models can help identify and catalogue these requirements, which can otherwise use up precious government resources. All this has to be done responsibly with rigorous testing because generative AI models have the potential to serve up hallucinations or incorrect information, Ho said. Still, there is a potential for use in areas where legal search and reform can be costly, he said.

Public Good

Ho said he has a personal connection to the racially restrictive documents. In 2015, when he purchased a home in Palo Alto, Calif., the historic deed said the “property shall not be used or occupied by any person of African, Japanese or Chinese or any Mongolian descent” except as a servant to a White person. Ho is a first-generation immigrant whose parents were originally from Hong Kong.

“It really was a compelling project to me, both because of the sheer challenge that it represents in terms of the number of deed records, but also the opportunity to actually really understand local history,” he said. “Even if unenforceable, those words can have powerful effects to signal the type of community that the home belongs to.”

Peter Henderson, a Princeton professor who co-authored the paper and researches AI machine learning and responsible use in the real-world, said the project showed the benefits of using open-source tools such as Mistral, meaning that it is freely available and can be changed and modified.

“If you’re trying to do this at scale, the open source really helps, and we can customize the models to do much better at these tasks,” Henderson said. “A proliferation of open source models has made it easier and more possible to take on these challenges affordably.”