Artificial intelligence may be quietly penalizing Americans not for who they are, but for where they live, according to USC researchers. Consider a young professional woman who has done everything right.

Before turning 30, she has built a promising career. She earns a good salary, pays her bills on time and maintains an excellent credit score. Yet when she shops for car insurance, companies quote her higher premiums than friends living just a few ZIP codes away. When she begins looking for a home, mortgage lenders offer less favorable rates than borrowers with nearly identical financial profiles.

She feels as though an invisible disadvantage follows her. She’s right. The problem isn’t her income. It’s where she lives.

“Some people lose money that they are unfairly paying toward higher insurance premiums. They lose access to certain services that can make their lives easier” because of spatial biases in the outcomes in AI decision-making models, said Nripsuta “Ani” Saxena, a doctoral student in the Thomas Lord Department of Computer Science in the USC Viterbi School of Engineering and the USC Stevens School of Computing and AI.

Saxena is the lead author of “Spatial Fairness: Foundations, Pitfalls, and a Path Forward,” a paper exploring how artificial intelligence-based decision making systems can unintentionally discriminate based on individuals’ residences or locations, even when attempts are made to account for legally protected attributes such as race and ethnicity.

She and her colleagues argue AI systems can infer far more from an address than most people realize. In many communities, residential patterns still reflect decades of housing segregation and unequal investment, allowing AI systems to produce unequal outcomes even when race and ethnicity are not considered by models directly.

The paper was co-authored by Abigail Horn, Saxena’s  co-supervisor, a research assistant professor in the Daniel J. Epstein Department of Industrial & Systems Engineering and a research computer scientist at the USC Information Sciences Institute; and Cyrus Shahabi, her doctoral advisor, the Helen N. and Emmett H. Jones Professor of Engineering, professor of computer science, electrical and computer engineering, and spatial sciences, and director of the Integrated Media Systems Center; and Wenbin Zhang, an assistant professor in the Knight Foundation School of Computing and Information Sciences at Florida International University. 

Together, the researchers argue that location has become one of artificial intelligence’s biggest blind spots.

Removing race isn’t enough

Evidence of location bias has surfaced repeatedly.

A ProPublica investigation found drivers living in minority neighborhoods often paid significantly higher car insurance premiums than drivers with similar risk profiles in predominantly white neighborhoods. Bloomberg reported Amazon’s Prime same-day delivery service excluded some predominantly Black neighborhoods, including an affluent one in Boston. Academic researchers have found that riders coming from neighborhoods of color and higher poverty levels are often charged higher fare prices.

None of those studies concluded companies intentionally discriminated based on factors such as race, ethnicity, and income. Instead, they pointed to a subtler problem: AI systems can learn historical patterns embedded in geographic data  – for example, higher crime rates in areas affected by redlining – and base decisions off of them in ways that disadvantage entire communities

At first glance, the solution seems obvious: Don’t let AI decision-making models consider race.

The reality is more complicated.

“Machine learning models designed to find patterns,” Horn said. Even if race is removed as a feature explicitly considered by a model, those patterns can still emerge through geography. A neighborhood can reflect decades of housing segregation, unequal investment and other historical factors, allowing AI to draw conclusions that resemble using race without ever explicitly considering it.

Yet location isn’t simply a stand-in for race. It also provides information companies legitimately need. Insurance companies need to know whether a home sits in a wildfire zone. Mortgage lenders need to understand local housing markets. Ride-sharing companies need to know where trips begin and end. Eliminating location altogether would make many AI systems far less useful.

“You can’t simply not use location,” Saxena said. “Location does give you legitimate information that is relevant to the decision-making task.” The challenge, she said, is making sure its correlation with legally protected characteristics “does not seep into your decision-making and color the outcomes.”

In other words, the goal isn’t to remove location from AI. It’s to ensure algorithms use the information that matters without allowing geography to become a proxy for characteristics they should never consider.

Otherwise, AI risks perpetuating yesterday’s inequalities using today’s maps.

Building fairer AI

Rather than proposing a quick technical fix, the paper calls for a different way of thinking about AI bias.

“The very first step is to be aware of it,” Saxena said. “You cannot fix something that you’re not aware of.”

That means asking tougher questions before AI systems are deployed. Could an algorithm unintentionally disadvantage people because of where they live? Has anyone tested for patterns that unfairly affect entire neighborhoods? Once a system is deployed, is anyone checking whether people in certain communities are consistently receiving worse outcomes than others? Some emerging work has begun to offer frameworks for auditing spatial bias in model outcomes, but more dedicated research is needed to develop models that are spatially fair by design.

The researchers also urge companies to involve legal experts and representatives from affected communities much earlier in the design process so potential problems are identified before algorithms begin shaping people’s lives.

For Saxena, the need to act is urgent.

“Spatial fairness, or lack thereof, truly affects the wallets, health, and daily lives of a significant chunk of this country every single month,” she said.

Published on September 2nd, 2026

Last updated on September 2nd, 2026