Key Takeaways
- At least 50 million Americans live under local laws that discriminate by race, gender, citizenship, age, or disability.
- An LLM-assisted pipeline scanned 9 million legal sections across 9,623 jurisdictions and surfaced roughly 10,000 suspect statutes.
- The AI framework accelerates legal discovery, but human review remains essential before reform can proceed.
Table of Contents
50 Million Americans Governed by Laws That Should Not Exist
Stanford HAI reports that at least 50 million Americans currently live under local laws that discriminate by race, gender, citizenship, age, or disability.
The finding comes from a Stanford Law School study that used LLM-assisted review to scan 9,623 local jurisdictions across the United States.
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.
Yasmine Mabene, a research fellow at Stanford RegLab and the study’s first author, delivered that assessment.
As detailed by Stanford HAI, the search surfaced roughly 10,000 suspect statutes buried inside approximately 9 million local legal sections.
Michigan still bars non-citizens from obtaining private investigator licenses.
Memphis carries a board of education provision requiring separate schools based on race, while Valparaiso, Florida, has not removed poll-tax language from its books.
The paper appears in the proceedings of the International Conference on Artificial Intelligence and Law.
No comparable local-code review had previously been attempted at this scale.
Inside the AI Pipeline That Scanned 9 Million Legal Sections
The research team first consolidated written laws from 9,623 local jurisdictions, covering an estimated 75 percent of the U.S. population.
They then built a multi-stage workflow that moved beyond simple keyword search.
The first pass flagged any statutes that referenced protected categories.
The second pass determined whether those laws treated people differently based on those characteristics.
Human reviewers filtered out legitimate differences, such as disability accommodations or gender-neutral language updates.
Finally, LLMs ranked the remaining laws as low, medium, or high priority for legal review.
- Non-citizen occupational bans: More than 2,000 statutes block non-citizens from professional licenses or occupations where citizenship should not matter, including running a bowling alley.
- Sex-based criminal statutes: Some local codes still define prostitution as a crime only for women.
- Racial separation clauses: A Georgia statute directs election officials to keep separate voter lists based on race, while a North Carolina law requires separate cemeteries for white and Black residents.
- Derogatory statutory language: More than 30,000 laws contain outdated or offensive terms for residents with intellectual disabilities.
The approach was not flawless.
It achieved over 90 percent accuracy in detecting differential legal treatment and correctly flagged 88 percent of provisions that human annotators considered high priority.
Human review remains essential before any legal action.
Why Local Legal Sludge Is an AI Scale Problem
The study is not simply an inventory of forgotten statutes.
Many of these laws remain active, not dormant.
A Massachusetts ruling recently denied a liquor license to a green-card-carrying Brazilian family based on national origin.
That outcome mirrors numerous laws surfaced by the pipeline.
Daniel Ho calls the local-code problem ‘policy sludge,’ a label that captures how easily obsolete statutes settle into municipal books when review capacity is scarce.
The AI framework changes the economics of statutory discovery.
Manual efforts have historically required commissions, law students, attorneys, and judges to spend enormous time compiling discriminatory laws.
Machine review removes the largest bottleneck: finding the statutory language that needs attention.
The authors caution that the pipeline does not detect laws that appear neutral on their face but create discriminatory outcomes in practice.
Nor does it assess whether an otherwise neutral statute is enforced in a discriminatory way.
For the AI sector, this is a clear demonstration of structured human-in-the-loop review at scale.
It shows how LLMs can be applied to high-stakes text analysis when the output is constrained, auditable, and paired with expert oversight.
The Reform Window AI Just Opened
The immediate payoff is practical: local governments and legal reform groups now have a usable map of discriminatory statutes, with enough precision to prioritize review and action.
For teams building AI-assisted review or compliance analysis at similar scale, programmatic SEO AI automation is how Andres SEO Expert approaches structured, high-volume text processing — contact Andres SEO Expert.
Frequently Asked Questions
What did Stanford’s AI legal review find about discriminatory local laws?
Stanford Law School used LLM-assisted review to scan 9,623 local jurisdictions and found that at least 50 million Americans live under overtly discriminatory local laws violating constitutional protections based on race, gender, citizenship, age, and disability. Roughly 10,000 suspect statutes were found among approximately 9 million local legal sections.
How did the AI pipeline scan millions of local legal sections?
The team consolidated written laws from 9,623 jurisdictions covering an estimated 75 percent of the U.S. population. They then built a multi-stage workflow: the first pass flagged statutes referencing protected categories, the second pass determined whether those laws treated people differently, human reviewers filtered out legitimate differences, and finally LLMs ranked the remaining laws as low, medium, or high priority for legal review.
What types of discriminatory local laws were uncovered?
The study found more than 2,000 non-citizen occupational bans, sex-based criminal statutes such as defining prostitution only for women, racial separation clauses like separate voter lists or cemeteries, and over 30,000 laws containing outdated or offensive terms for residents with intellectual disabilities.
Are these discriminatory local laws still actively enforced?
Yes, many of these laws remain active, not dormant. For example, a Massachusetts ruling denied a liquor license to a green-card-carrying Brazilian family based on national origin, which mirrors numerous laws surfaced by the AI pipeline.
What is ‘policy sludge’ in the context of local laws?
Daniel Ho uses the label ‘policy sludge’ to describe how easily obsolete statutes settle into municipal books when review capacity is scarce. The AI framework changes the economics of statutory discovery by removing the largest bottleneck: finding statutory language that needs attention.
What are the limitations of the AI legal review approach?
The pipeline does not detect laws that appear neutral on their face but create discriminatory outcomes in practice, nor does it assess whether an otherwise neutral statute is enforced in a discriminatory way. It achieved over 90 percent accuracy in detecting differential treatment but human review remains essential before any legal action.
