A new Stanford study used AI to scan over 9 million legal sections from 9,623 local jurisdictions and found roughly 10,000 laws that likely discriminate based on race, gender, citizenship, age, or disability. The research, published in the proceedings of the International Conference on Artificial Intelligence and Law (ICAIL), estimates that at least 50 million Americans live under local codes that still contain overtly unconstitutional provisions.
"We estimate that at least 50 million Americans are living in cities, counties, or towns governed by overtly discriminatory laws that violate constitutional protections based on race, gender, citizenship, age, and disability," said Yasmine Mabene, a research fellow at the Stanford RegLab and first author of the paper. "The AI pipeline we developed can find these laws and help get them off the books."
The scale of the problem
Local laws have long escaped the scrutiny applied to federal and state statutes. While a 2019 Virginia commission required law students from three universities, governor's office staff, practicing attorneys, and a sitting judge just to compile discriminatory laws from a single state over a 60-year period, no comparable effort had ever tackled local codes nationwide. The reason is simple arithmetic. "Finding the laws on the local books would have been incredibly time consuming without the help of AI," Mabene said.
Daniel Ho, professor at Stanford Law School, director of the RegLab, and the study's senior author, called these lingering statutes "policy sludge." He said local codes remain particularly "sticky" due to lack of local resources. "While many manual efforts have focused on reforming federal and state laws, local laws stay on the books because no one has had the time to look and take action."
How the AI pipeline worked
The research team designed a multi-stage workflow that went beyond simple keyword searches. First, an LLM flagged any statute referencing a protected category - race, religion, sex, gender, sexual orientation, marital status, national origin, genetic information, age, or disability. Second, it determined whether those laws treated people differently based on those characteristics. Human reviewers then filtered out acceptable differences, such as disability accommodations or gender-neutral language updates, and the LLM ranked remaining laws as low, medium, or high priority for legal experts to review.
"This was more than simple word search for discriminatory language," Mabene said. "We had to design an approach that could work through the same legal questions a human reviewer could ask, distinguishing between laws that actually discriminate versus laws that simply mention protected groups by name." The method achieved over 90 percent accuracy in finding differential legal treatment and correctly flagged 88 percent of provisions human annotators also deemed high priority. For legal professionals interested in how AI is reshaping document review and research, AI for Legal Professionals Courses cover the tools and workflows now entering practice.
What the search uncovered
The pipeline surfaced more than 2,000 laws barring non-citizens from obtaining professional licenses or working in occupations where citizenship should not matter - including running a bowling alley. It found laws that make prostitution a crime only for women, a Georgia statute directing clerks to keep "separate lists of white and colored voters," and a North Carolina law requiring "separate cemeteries for white and black."
Beyond explicitly discriminatory provisions, the AI flagged over 30,000 laws containing offensive language. One Ohio town's housing statute refers to intellectually disabled residents as "feeble-minded, insane, lunatics, imbeciles or idiots." The researchers published a companion website where users can explore the detected provisions. These findings show that the problem is not merely historical residue. A recent Massachusetts ruling denying a liquor license to a green-card-holding Brazilian family based on national origin mirrors laws the pipeline surfaced, confirming that many of these statutes are still enforced.
The researchers acknowledge limitations. The method does not identify facially neutral laws with discriminatory effects, nor does it address discriminatory enforcement of otherwise neutral laws. Human review remains essential. Still, for paralegals and legal researchers managing large-scale document review, the workflow offers a model for what AI-assisted statutory analysis can achieve. AI for Paralegals Learning Path explores similar applications in legal research automation.
Why this matters for legal professionals
For lawyers, paralegals, and compliance officers, the study demonstrates that AI can now handle statutory review at a scale no human team could match - but it does not replace human judgment. The pipeline's 88 percent accuracy on high-priority flags is strong, yet the remaining 12 percent means legal expertise is still the final safeguard. The takeaway is practical: AI can surface the needles in a haystack of millions of legal texts, but a licensed professional must still decide which needles actually draw blood. Law firms and government legal departments that learn to pair LLM-driven review with attorney oversight will have a structural advantage in due diligence, regulatory compliance, and civil rights litigation.
Your membership also unlocks:
