Ethical AI Practices & Data Governance Flashcards

A bank's loan model is audited and found to have higher false positive rates for minority applicants (incorrectly flagging them as high-risk). Which remediation technique directly targets this disparity?

Answer: Applying reweighting or post-processing threshold adjustments per group

Reweighting training samples or applying group-specific decision thresholds via post-processing are standard fairness interventions that directly correct disparate false positive rates.

Which data governance role is PRIMARILY responsible for defining how a specific dataset may be used, ensuring compliance with policies and regulations?

Answer: Data steward

A data steward is responsible for defining and enforcing data usage policies, access rules, and compliance requirements for specific datasets within an organization.

The concept of 'algorithmic accountability' most directly implies that:

Answer: Organizations must be able to explain, audit, and accept responsibility for algorithmic decisions

Algorithmic accountability means organizations can justify their AI decisions, submit to audits, and bear responsibility for harms caused by their systems.

A machine learning team is designing a system to predict recidivism risk for parole decisions. The MOST ethically critical design requirement is:

Answer: Ensuring the system provides decision support, not autonomous binding decisions

In high-stakes criminal justice settings, ethical AI design mandates human decision authority; the model must be a tool supporting, not replacing, human judgment.

Under the principle of 'privacy by design,' privacy protections should be:

Answer: Embedded proactively into the system architecture from the start

Privacy by design (Cavoukian) requires that privacy controls be built into system architecture from inception rather than retrofitted post-deployment.

A model's calibration is assessed using reliability diagrams. A poorly calibrated model that consistently predicts 80% confidence but achieves only 60% actual accuracy is said to be:

Answer: Overconfident

Overconfidence means predicted probabilities are systematically higher than the actual observed frequencies, which reliability diagrams reveal as the calibration curve falling below the diagonal.

A governance team requires all third-party ML vendors to complete a 'Data Processing Agreement' (DPA) before accessing customer data. This requirement is MOST directly mandated by:

Answer: GDPR Article 28

GDPR Article 28 mandates that controllers use only processors providing sufficient guarantees and must enter into a binding Data Processing Agreement with each processor.