Historical Bias in Machine Learning

Historical bias is a source of harm that arises before or during model development because the data encode inequities that already exist. Suresh and Guttag’s machine-learning lifecycle framework distinguishes this from representation bias, measurement choices, aggregation, learning, evaluation, and deployment harms. A dataset can represent its source population accurately and still preserve unfair social structures. More data from the same process may make the pattern more statistically stable without making it fair. The mechanism depends on the target and data-generating process. If a hiring system predicts prior hiring decisions, it may treat past choices as a definition of merit. If a credit model uses variables shaped by segregation or exclusion, those variables can carry historical differences into future decisions even when race is omitted. In health care, Obermeyer and colleagues found that a widely used population-health algorithm used health-care cost as a proxy for need; because less money had historically been spent on Black patients with comparable illness, the score understated their need. The result was a target-design problem tied to structural differences, not simply an incorrectly measured feature. Historical bias differs from a data-coverage problem: adding more records of the same decisions may not address a discriminatory target. It also differs from label bias caused by annotator judgments, although both can be present together. Teams should ask what outcome the model is trained to predict, who had access to the opportunity in the past, and which institutional choices produced the labels. Mitigation starts by questioning the target and decision process. Use outcomes closer to the intended goal, examine whether historic labels are suitable, test subgroup impacts, and consider whether the institution should change the process rather than automate it. A fairness metric alone cannot decide what outcome is legitimate. Document the historical context, affected groups, and remaining uncertainty before deployment.