Bias, fairness and accountability in using models
Where it's dangerous
Any use that affects people's opportunities: hiring, credit, pricing, service priority, moderation. There the error isn't inconvenience but real harm, and sometimes legal exposure too.
What to do in practice
Measure performance across segments — not just an overall average — and check whether a particular group systematically gets worse results.
Keep a human in the final decision, and give them the information to decide differently and not just approve.
Document: what the system decides, on what basis, and how to appeal. Explainability is a practical requirement, not philosophy.
Transparency to users
People need to know they're interacting with an automated system, and what happens to their information. In regulated domains it's worth clarifying the specific requirements with someone qualified.
Going deeper
Build a dedicated fairness test set with parallel cases that differ only in a detail that shouldn't matter, and verify the result is identical. Run it as a regression on every model or prompt change — a version change can reopen a gap that was closed.