AI Ethics, Transparency, and Explainability
Free CSA Trusted AI Safety Expert (TAISE) practice — 6 questions on AI Ethics, Transparency, and Explainability, with explanations. No sign-up. Full 12-question mixed test →
A consumer credit-scoring model is already embedded in a regulated loan origination pipeline with a frozen model version approved by auditors. Post-deployment monitoring reveals a disparate impact ratio of 0.72 for a protected demographic group (below the 0.80 four-fifths threshold). Regulatory constraints prohibit retraining or redeploying a new model version for at least two quarters. Which bias mitigation approach BEST addresses the disparity within these constraints?
Equalized-odds post-processing operates purely on the already-produced model outputs (scores/decisions), recalibrating group-specific thresholds to equalize error rates without touching model weights or requiring retraining or redeployment, making it the only option compatible with a frozen, regulator-approved model.
A regulator demands a pixel-level attribution map showing which regions of a chest X-ray most influenced a deep convolutional network's diagnosis, so radiologists can visually verify the model focused on clinically relevant regions rather than spurious artifacts. Which explainability technique is BEST suited to this requirement?
Integrated Gradients is a gradient-based attribution method designed for differentiable models like CNNs; it produces fine-grained, pixel-level attribution maps by integrating gradients along a path from a baseline to the input, directly meeting the requirement for precise visual attribution.
An LLM-based hiring screening tool rejects a candidate, who formally challenges the decision alleging discriminatory bias. According to TAISE accountability principles, what should the AI governance team do FIRST before taking any corrective action?
Accountability frameworks require traceability before remediation: reviewing documented model cards, decision logs, and the audit trail allows the governance team to verify what data and logic actually drove the specific decision, which is the evidentiary basis for any further action and is required for defensible, auditable accountability.
An organization is designing the human-in-the-loop configuration for an automated content moderation system that must operate at scale while still meeting responsible-AI oversight requirements. Which configuration BEST balances oversight with operational scale?
Confidence-threshold-based routing is the standard human-oversight pattern that preserves meaningful human review of uncertain, high-risk, or borderline cases while allowing high-confidence decisions to scale automatically, satisfying both the operational and oversight requirements.
Which statement correctly differentiates 'interpretability' from 'explainability' as used in TAISE terminology?
TAISE distinguishes interpretability as an inherent property of simple, transparent model architectures where the decision logic is directly readable, from explainability as the set of post-hoc methods (SHAP, LIME, attribution maps, counterfactuals) applied after the fact to approximate or expose the reasoning of complex black-box models.
During intersectional bias testing of a facial recognition system, auditors find the model has similar overall accuracy across demographic groups, but one subgroup experiences a significantly higher false positive rate than others, leading to disproportionate false accusations for that group. Which fairness metric is MOST directly relevant to detecting and addressing this specific disparity?
Equalized Odds requires that both the true positive rate and the false positive rate be equal across groups; it directly captures and would flag the disparity described, since false positive rate parity is an explicit component of this metric, unlike metrics focused only on overall selection rates or precision.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →