Existing laws already fight AI housing discrimination—new state AI bills increase confusion

In 2021, Mary Louis of Massachusetts had her application for an apartment she hoped to rent rejected because a computer algorithm flagged her as a financial risk. The following year, she and co-plaintiff Monica Douglas filed suit at the head of a class of 400 “low-income, minority housing voucher holders,” alleging they were “effectively blackballed from rental housing by Defendant SafeRent Solutions, LLC based on credit histories and other information which bears little to no relationship to the risk that their rent will not be paid.”

SafeRent settled the suit out of court, agreeing to pay over $2.2 million and modify certain features of the scoring algorithm it offered to property owners to evaluate prospective tenants. The case received little news attention at the time of its filing in 2022, but that changed when its settlement was approved in late 2024.

“She didn’t get an apartment because of an AI-generated score — and sued to help others avoid the same fate,” proclaimed a Guardian headline in December. Associated Press coverage of the settlement explained that:

While such lawsuits might be relatively new, the use of algorithms or artificial intelligence programs to screen or score Americans isn’t. For years, AI has been furtively helping make consequential decisions for U.S. residents.

When a person submits a job application, applies for a home loan, or seeks particular medical care, there’s a chance that an AI system or algorithm is scoring or assessing them, just as it did with Louis. Those AI systems, however, are largely unregulated, even though some have been found to discriminate.

Persistent discriminatory outcomes, even without direct intent to discriminate, are unfortunately not a new phenomenon in housing markets. Neither are computer algorithms, which have been commonly used throughout housing markets for decades and are subject to existing anti-discrimination laws. What is new is calling these algorithms “artificial intelligence” or “AI.” At the time of the alleged discrimination in 2021, a computer program that generated a financial score would not commonly have been called AI. The term does not appear a single time in the 43-page complaint filed in 2022.

These years coincide with the widespread adoption of AI chatbots and their solidification in public awareness. There’s no indication that SafeRent’s algorithm used any of the new technology employed by these chatbots, such as large language models. If, in 2021, its designers had searched for an of-the-moment buzzword to describe their product, they would more likely have landed on “big data” or “machine learning” than AI. This had changed by 2024, and “AI” is found in almost every headline covering the settlement.

Legally speaking, this does not matter. Whether SafeRent’s algorithm employed a large language model or older programming technology does not impact whether it violated discrimination law. In general, there is no clear line between what we call AI today and many of the algorithms we have used for years without that moniker. It is not surprising that the big tent of what we call AI continues to grow. Companies want to use the term for branding purposes, just as the media does for headlines.

However, this rapid renaming of many algorithms as “AI” matters in the realm of policy. Both regulators and the public are willing to consider bolder and more sweeping regulation of AI than would be likely with more incremental technological change. Mainstream policy proposals for regulating AI have run the gamut from a somewhat fantastical moratorium on AI innovation itself to a moratorium on regulation at the state level. The technologies placed in the “AI” bucket may end up on a very different regulatory trajectory than those that don’t.

The combination of technology-driven anxiety and media hype is behind a wave of state legislation aimed at protecting housing markets from AI-driven discrimination. These bills create costly disclosure requirements and rules, which are necessarily based on speculation about what forms potential threats from the still-evolving technology might take. In Virginia, Gov. Glenn Youngkin vetoed House Bill 2094 for these reasons, stating that the bill’s “rigid framework fails to account for the rapidly evolving and fast-moving nature of the AI industry and puts an especially onerous burden on smaller firms and startups that lack large legal compliance departments.” A similar bill was signed into law by Colorado Gov. Jared Polis, who, in an unenthusiastic signing statement, expressed support for a federal moratorium on such state efforts.

Misguided AI regulatory efforts risk limiting innovation and sowing misunderstanding in many markets. However, the long, complex, and still-pertinent history of U.S. housing discrimination makes the potential damage even greater. Cases like SafeRent are sometimes referred to as “digital redlining,” reflecting how discriminatory outcomes can inadvertently emerge from data-driven algorithms. However, this term is somewhat misleading, as it harkens back to the mid-20th century when housing discrimination was overt and government-led.

Between 1935 and 1940, the federal Home Owners’ Loan Corporation (HOLC) created maps and neighborhood taxonomies to help guide the lending decisions of banks and mortgage lenders. HOLC gave each neighborhood a grade of A through D, with the neighborhoods receiving a grade of D notoriously marked in red and designated “hazardous.”

HOLC did not attempt to hide its discriminatory aims, docking numerous city points for neighborhoods where the number of immigrants and black Americans passed certain thresholds. Low-income minority communities were systematically denied access to credit and financial capital, essential for robust economic activity. HOLC’s condemnation of these neighborhoods was based on nothing more than racism and xenophobia, but the agency’s vast powers meant its pronouncements did significant harm to some neighborhoods over time. The lines drawn on maps around “hazardous” neighborhoods gave the practice its name—redlining.

In contrast, digital redlining refers to modern cases where computers use data or information reflecting disparities stemming from these once-intentional and widespread practices. Researchers see the persistent impact of 20th-century redlining and other forms of housing discrimination on current financial and housing market outcomes.

There are no easy answers to leveling the playing field or quantifying the impacts on minority homeowners and renters today. But state legislatures should take note that the class action lawsuit and settlement were brought under existing anti-discrimination law. As AI does not represent a break in the algorithmic tools being used, merely incremental improvements, there is no reason to suspect efforts like the Virginia and Colorado bills would succeed where other government efforts have failed.

As the AI revolution brings new algorithmic tools to many markets—and rebrands existing tools as “AI” in others—there exist opportunities both for learning and greater honesty in debates about housing discrimination. Different types of discrimination must be clearly distinguished. Twentieth-century redlining was intentional and government-led. Digital redlining, in contrast, lacks direct “bad guys” to deter or prosecute. Attempting to curb bad results by going after technology is tempting for some, but it has no track record of working.

The post Existing laws already fight AI housing discrimination—new state AI bills increase confusion appeared first on Reason Foundation.

Source: https://reason.org/commentary/existing-laws-already-fight-ai-housing-discrimination-new-state-ai-bills-increase-confusion/