When Your Tool Treats Groups Differently

So you ran the fairness check and found a gap: your tool treats one group differently from another. Don't panic — and don't hide it. Finding the gap is the system working. Here's the practical playbook.

First, quantify it. How big is the difference, and on what decisions? A small gap on low-stakes recommendations is a different conversation than a large gap on hiring or lending. Second, find the driver: which data or design choice is producing it? Proxies, unrepresentative training data, and miscalibrated thresholds are the usual suspects.

In plain English: measure it, find the cause, then choose your response with eyes open.

Your options: fix the data or model, narrow the tool's role (advisory instead of decisive), add meaningful human review for affected cases, or pause deployment for that use. What you can't do — the one option that's off the table — is document nothing and ship anyway.

Write it all down: the finding, the analysis, the decision, the sign-off. That record turns a scary discovery into evidence of responsible governance. Regulators and auditors don't expect perfection. They expect a process.

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