Ethical AI Decision Making

Decision Weights

Accuracy weight 0.50

Proxy-signal weight 0.50

Approval threshold 0.50

Outcome

Approval rate — Group A

–

Approval rate — Group B

–

Disparate-impact ratio (four-fifths rule)

–

How it works

Every candidate carries two features: a true qualification score (equally distributed in both groups) and a proxy signal — a stand-in for the kind of correlated-but-biased feature that creeps into real training data (e.g. a "culture fit" score, a zip-code-derived risk band). In this scene Group B's proxy signal is drawn from a lower distribution than Group A's, even though qualification is identical — exactly the kind of historical bias the article describes: the algorithm isn't malicious, it just learned a pattern baked into its data.

score = w_acc·qualification + w_speed·proxy
approve if  score ≥ threshold
disparate impact = min(rateA,rateB) / max(rateA,rateB)

Accuracy weight — how heavily the true qualification score counts toward the decision.
Proxy-signal weight — how heavily the biased proxy counts. Push this up and Group B's approval rate drops even though nothing about their real qualification changed.
Approval threshold — the cutoff score; raising it approves fewer candidates overall from both groups.
Bias correction — a simple reweighing pass that re-centers each group's proxy signal on the same mean before scoring, so the proxy stops silently encoding group membership.

The four-fifths (80%) rule is a common regulatory rule of thumb: if one group's selection rate falls below 80% of the highest group's rate, the outcome is flagged as evidence of disparate impact — visualized here by the ratio meter turning red.

Approved Denied Group A floor Group B floor

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.