What are racially restrictive covenants?
Racially restrictive covenants are discriminatory clauses that prohibited the purchase, lease, or occupation of land based on race. Racial covenants originated in the mid-19th century, but became more prevalent after the Supreme Court held racial zoning unconstitutional in Buchanan v. Warley (1917). Instead, homeowners turned to private transactions and racial covenants as a way to protect property values and maintain racial homogeneity within their communities. Although declared unenforceable by the United States Supreme Court in Shelley v. Kraemer (1948) and illegal under the Fair Housing Act (1968), such covenants continue to exist in the pages of real property records across the nation.
What is California's AB 1466?
California enacted legislation in 2021 (AB 1466) that mandates that counties redact racially restrictive deed covenants. Each county must develop an implementation plan to identify and redact such records. For a county like Santa Clara, this involves a substantial effort to potentially examine 24 million deed documents, spanning 84 million pages, dating back to the 1850s.
Prior to AB 1466, individual homeowners could petition for legal review of covenants, but as in other jurisdictions putting the onus on homeowners, few individuals did so.
In other areas, groups have crowdsourced the identification of racial covenants, with thousands of volunteers reading over deed records (see National Covenants Research Coalition). Such initiatives are great for involving community members, but infeasible in jurisdictions lacking that kind of volunteer base.
How did Stanford RegLab and Santa Clara County collaborate?
Stanford RegLab and the Santa Clara County Clerk-Recorder's Office collaborated to explore the use of machine learning to solve this challenging search task. RegLab amassed a training dataset drawn from eight counties across the United States, developed and fine-tuned a large language model to detect racial covenants, and created a registry to map these covenants in the county. The collaboration saved some 86,500 person hours that would have been required for manual review of records and the model will be made available for free for any jurisdiction and research team to use.
With 57 other counties in California working on the implementation of AB 1466 and some dozen states enacting reform around racial covenants, the model can dramatically lower the cost of identifying such provisions.
Why does identifying racial covenants matter?
California mandated this kind of review because of the significant stigmatic harm that can come from racial covenants in the public record, a large number of which employ racial epithets. Professors Richard Brooks and Carol Rose also argue in a leading book that the effect of racial covenants persists due to what they signal. In 2002, for instance, a man in Richmond VA refused to sell his home to an African American woman, pointing to the racial covenant. Even if unenforceable, those words can serve as a signal of the type of community.
A historical accounting also informs prospective efforts at fair housing. The study shows that just a small number of developers were responsible for close to a third of racial covenants. Joseph Eichler, who famously built ~2700 homes in Palo Alto, however, firmly resisted such covenants, and this shows the power of a small number of actors in influencing integration and fair housing.
What other areas of legal reform are like this?
The study demonstrates the power of language models to assist in legal reform. Reform efforts often run into the complexity of language littered across codes, regulations, and records. The project demonstrates a path to using technology to aid in such efforts. To provide an example other than racial covenants, consider Congressionally mandated reports that agencies have to complete. The number has ballooned over time, but it's difficult to count such reports across U.S. Code, let alone determine which ones are ripe for sunsetting. The power of language models can help identify, catalog, and streamline otherwise opaque requirements that can waste government resources. As the team demonstrated with racial covenants, this can be done with rigorous evaluation, testing, adaptation and legal oversight.
What are the implications for technology and government?
Academic-agency partnerships provide an alternative path for responsible AI integration. The project demonstrates the substantial value of open innovation. The cost of the open model is less than 2% the cost of a comparable proprietary model and the model will be freely available for other jurisdictions to use.
This is great! How can I do this too?
If you are a jurisdiction or a party with access to deed records and are interested in carrying out a similar project with our system, please feel free to reach out to us at reglab [at] law [dot] stanford [dot] edu. We are not able to respond to all inbounds, but it is particularly helpful if you can let us know who you are, what jurisdiction you are interested in, and what deed record data is available. We've also made available our trained model for racial covenant detection on HuggingFace.