Datasets & Benchmarks · 1996
Adult Income / Census
A slice of the 1994 US census became the dataset that taught machine learning about fairness.
In the mid-1990s, Barry Becker and Ronny Kohavi extracted around 48,000 records from the 1994 US census database and donated them to the UC Irvine Machine Learning Repository in 1996. The task they defined was simple: predict whether a person earns more than 50,000 dollars a year from attributes like age, education, occupation, marital status, hours worked, sex and race.
Kohavi used the data that same year in research on hybrid decision-tree and naive Bayes classifiers, and the dataset, known as Adult or Census Income, quickly became a standard benchmark. For two decades, almost every new tabular classifier was measured against it.
Its second life proved more important than its first. Because the data pairs income labels with sensitive attributes like sex and race, it became the default testbed for algorithmic fairness research. Landmark work on demographic parity, equalized odds and bias mitigation almost always reported results on Adult.
That ubiquity eventually drew criticism. The records reflect the 1994 economy, the 50,000 dollar threshold is arbitrary and interacts oddly with inflation, and researchers argued the dataset had become a ritual rather than a realistic test. In 2021 a team including Frances Ding and Moritz Hardt released modern replacements built from newer census releases.
Even so, Adult remains a touchstone. It marks the moment machine learning started grappling with a question that now sits at the center of enterprise AI: not just whether a model is accurate, but for whom.
From history to production
We turn these ideas into working systems
The same techniques, shipped into your stack with evals, observability, and measurable ROI.