(Illustration by iStock/Infadel)
Credit data represents one of the earliest instantiations of big data in American life, with credit bureaus serving as the nation’s original data brokers. What began as a narrow system for evaluating financial risk has since expanded far beyond the boundaries of financial history as corporations across industries deploy algorithms of every conceivable kind to sift through personal data and determine access to opportunity.
Barbara Kiviat, an assistant professor of sociology at Columbia University; Sara Sternberg Greene, a professor of law at Duke University Law School; and Hesu Yoon, an assistant professor of sociology at CREST-ENSAE Paris, teamed up to study one consequential corner of the big-data economy: tenant screening.
Landlords, property managers, and real estate companies now rely on an expanding arsenal of personal data—including credit scores, criminal records, and eviction histories—to assess would-be renters. Long before any human conversation takes place, algorithms shape tenant screening, helping decide whether someone is approved or denied. For many landlords and property managers, this data-driven process has an appealing aura of objectivity, consistency, and fairness. But does it leave no room for human discretion?
The researchers uncovered a revealing story not in the rejections these systems produced, but in the exceptions: cases where an applicant’s record raised red flags but was approved anyway. In studying these exceptions, they found that human judgment persisted behind the facade of a neutral algorithm. “One of the paper’s big findings is that landlords and property managers are making exceptions, but they’re what we call systematic exceptions,” Kiviat says. “It’s not for one person at a time, but instead for one category of person at a time.”
The researchers distinguish between two modes of tenant screening: judgmental, or case-by-case evaluation, and algorithmic, where rules set in advance apply to all applicants. In both cases, holding out for perfect records isn’t viable for property owners seeking to avoid empty units and thus financial losses. In both types of tenant screening, landlords made exceptions for similar reasons, drawing on commonly held narratives to differentiate applicants whose blemished records signal real risk from those shaped by circumstances beyond their control. The 2008 financial crisis offers a telling example: Once foreclosures were widely understood as the result of predatory lending instead of personal failure, landlords judged them more leniently. The narrative shifted and exceptions followed.
“Just because organizations are making data-fed decisions with rules and with algorithms, that doesn’t mean they aren’t making exceptions,” Kiviat says. “It doesn’t mean they’re not treating some people with bad records as special and others as not.”
To understand how those exceptions work, the researchers conducted in-depth interviews with 88 individuals involved in leasing decisions in two metro areas: Durham, North Carolina, and San Jose, California. To build their sample, they used random selection from rental-housing websites and recruited harder-to-find respondents, including small owner-operators and senior executives. The key difference was in the kind of exception each system could accommodate. Judgment-based decisions allowed for idiosyncratic exceptions shaped by the particulars of a case, while algorithmic systems required exceptions to be specified in advance, limiting them to predictable categories. Algorithmic screening did not remove human judgment but instead standardized it.
“This research is a really valuable intervention into how we understand how rules work in the digital age,” says Karen Levy, an associate professor of sociology at Cornell University. “The conventional wisdom we often hear is that one problem with algorithmic systems is that they foreclose this kind of thick, detailed consideration of specific circumstances. But this research shows why that’s not really an accurate representation of how algorithmic rules work. Algorithmic rules still afford exceptions, but they do it in a different way: Rather than being overridable in the moment based on emergent circumstances, exceptions are baked into the rule ex ante.”
The implications extend well beyond housing, Kiviat says. Algorithmic decision-making governs access to jobs, insurance, and loans, and the same questions apply in each domain. Who benefits from exceptions that are built in? Who is left out because their circumstances were too idiosyncratic, too rare, or too hard for a room of executives to anticipate?
“This is the new stratification regime,” Kiviat
says. “This is how stratification now works. Class, race, and gender have not
ceased to matter, but layered on top is an expanding infrastructure of
data-driven gatekeeping that is becoming just as determinative of who gets what
in American life.”
Find the full study: Barbara Kiviat, Sara Sternberg Greene, and Hesu Yoon, “Exceptions in the Algorithmic Age: Evidence from the Case of Tenant Screening,” American Journal of Sociology, vol. 131, no. 4, January 2026.
Read more stories by Daniela Blei.