Knowledge MaxxingHistorical bias
Historical bias occurs when a dataset accurately reflects a real past pattern of unfair treatment, so a model trained on it learns to continue that same unfairness going forward.
Prerequisite: Proxy variable
What matters
Unlike other forms of bias caused by flawed data collection, historical bias can appear even in perfectly accurate data, since the past pattern itself was the unfair one.
Example
A hiring model trained on a company's past hiring data may learn to disadvantage a group that was historically underhired, even if the data accurately reflects what actually happened.
Common misconception
Historical bias is not solved simply by ensuring the data is accurate; the data can be perfectly accurate and still encode a pattern that shouldn't be continued.
Recall question
Why can perfectly accurate data still produce a model with historical bias?