Two demographic groups (blue and amber particles) flow left to right through a small feed-forward network. Each output node lights green (approve) or red (deny). The panel lets you tune four classic AI-ethics levers and watch the metrics respond immediately.
Demographic parity gap: Δ = |P(approve | group A) − P(approve | group B)|
Differential privacy: noise ~ Laplace(1/ε), smaller ε → more noise → stronger privacy, lower accuracy
Fairness constraint: reweights the decision boundary toward Δ → 0, trading off raw accuracy
Human oversight: a reviewer node re-checks denials and reverses those driven mainly by group membership
- Fairness constraint — pulls the two groups' approval rates together (lowers Δ); pushing it to 100% removes the bias term entirely but costs some raw accuracy.
- Differential-privacy noise — adds Laplace noise to each input before it reaches the network (visualized as jitter); more noise protects individuals but blurs the decision boundary.
- Transparency — "Explainable" lights up the strongest weighted edges on each pass, approximating a feature-importance / SHAP-style explanation of why a node fired.
- Human oversight — toggles a reviewer node that intercepts denials correlated with group membership rather than input features, and flips them to approve.
These four levers mirror the trade-offs discussed in AI-ethics practice: fairness-by-design, privacy-by-design, explainable AI (XAI) and human-in-the-loop accountability rarely optimize independently — improving one usually costs a little of another.