Ethics, Bias and Responsible Deployment

Models copy the past. If hiring data is biased, the model will be biased. Ethics is not an afterthought, it is part of building a system that works for everyone.

You need to check performance by group, not just overall. A model can be 95% accurate overall but 50% for a minority group. This is how real products fail and harm.

Responsible deployment means monitoring, not just launching. Data drifts, and a model that was good last month may be bad today. Plan to watch it.

Check bias

Slice metrics by group. Compare precision/recall for different demographics or regions. If a group is worse, collect more data or adjust thresholds.

from sklearn.metrics import classification_report
print(classification_report(y_test, pred, target_names=["groupA","groupB"]))
# Check per-group

Deployment and monitoring

Deploy with a canary, log predictions, and monitor drift. Have a rollback plan.

# Monitor: track p(y) over time
# If drift > threshold, retrain
# Keep human in the loop for high-stakes decisions

Syntax

# Check per-group metrics
for group in ["A","B"]:
    mask = df["group"]==group
    print(group, accuracy_score(y_test[mask], pred[mask]))

# Keep human in the loop for high-stakes decisions
# Log predictions and allow appeal