Data Ethics

Principles

• Fairness — no discriminatory outcome.
• Accountability — clear ownership.
• Transparency — explainable decision.
• Privacy — respect user data.
• Robustness — model + system reliable.
• Human oversight — escalation path.

Common bias

• Historical bias — past data reflect past inequality.
• Sampling bias — training data not representative.
• Aggregation bias — same model for different groups.
• Confirmation bias — analyst find what they expect.

Mitigation

• Diverse team — different perspective surface.
• Fairness metric (equal opportunity, demographic parity).
• Adversarial testing — try to break model.
• Human-in-the-loop for high-stakes.

Regulatory

• EU AI Act 2024 — risk-based, high-risk AI (credit, hiring) regulated.
• VN — chưa có AI specific law; PDPL apply.
• Sectoral — SBV credit scoring oversight.

Examples (negative)

• Amazon hiring AI bias against women (sunset 2018).
• COMPAS recidivism — racial bias.
• Apple Card credit limit gender disparity.

Internal governance

• AI ethics committee (legal + tech + business + ethicist).
• Model card — document model + limitation.
• Quarterly bias audit.
• Sunset criteria — when retire?

VN sensitive

• Religion, ethnicity data — extra care.
• Geographic bias — Hà Nội vs Hồ Chí Minh vs nông thôn.
• Age + gender — common in Vietnamese culture, but legal exposure.