Your Loan Algorithm Still Has to Be Fair

Here's a sentence worth taping to your monitor: the law doesn't care that it was the algorithm. America's fair lending laws — the Equal Credit Opportunity Act and its siblings — prohibit discrimination in credit decisions. They were written before machine learning existed, and they apply to it anyway.

That means if your model denies loans, you need to know why. Not "the model said so" — the actual reasons, in terms a regulator and a rejected applicant can understand. You need to test for disparate impact: does the model deny one group more often than another, even without using protected characteristics directly? Zip codes and shopping habits can be proxies.

The fix is the same discipline as everywhere else: test before deploying, monitor after, document everything, and keep humans who can explain and override decisions.

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