Deconstruct real-world biased loops. Identify proxies, classify bias types, and locate the human–algorithm boundary.
OmniHire Resume Ranker was trained on 10 years of mostly male hiring data. Which variable is the poisoned proxy that systematically down-ranks female candidates?
MediCare Priority Predictor uses past healthcare expenditure as a proxy for illness severity. Lower-income patients appear “low risk.” What kind of bias is this?
Where does human responsibility end and algorithmic autonomy begin? Drag statements into the correct zone.
A policing algorithm sends more patrols to neighborhoods with high past arrest rates. More arrests occur → the model reinforces the pattern. What is this?
Build the four non-negotiable pillars that will govern every AI system you design.
When should acceptance rates be forced equal across groups, and when should pure individual merit decide?
A rejected loan applicant demands an explanation. Which form meets a strong transparency requirement?
Design the rule that forces Pixi to halt and escalate to a human reviewer.
How should an affected user be able to appeal an algorithmic decision?
Architect EquiLoan — an auditable, fair micro-business loan system from data to explanation.
Strip protected attributes and detect subtle proxies. Drag dangerous fields into the “Strip / Monitor” zone.
Choose and defend the quantitative fairness benchmark for EquiLoan.
Build a live, user-facing explanation. Adjust the controls — the card updates instantly.
Pixi asks: “If a protected demographic’s average score drops by 5 %, what does your alerting pipeline do?”