The Bias You Encoded Without Knowing: Proxy Variables

Many teams think they've handled AI bias by removing sensitive fields — no race, no gender, no age in the data. Then the system discriminates anyway. How?

Through proxies. A proxy variable is an innocent-looking field that quietly stands in for the thing you removed. Zip codes can stand in for race and income. School names can stand in for class background. Employment gaps can stand in for caregiving — which correlates with gender. The model doesn't know what a zip code "means." It just finds that certain zip codes predict the outcome, and acts on it.

In plain English: deleting the label doesn't delete the pattern. The bias just puts on a disguise.

This is why "we don't collect that data" is not a fairness strategy. The fix is outcome testing: check results by group regardless of what fields you removed. If the outcomes skew, something in your data is acting as a proxy — and you find it by looking at results, not inputs.

Fairness can't be achieved by deletion. It has to be verified by measurement.

I explain frameworks; I don't give legal counsel.