Disparate-impact liability can arise from a facially: Under nondiscrimination law, why can this still create
Neutral models can still violate nondiscrimination law through disparate impact when they produce significantly unequal outcomes.
The question
A bank deploys an AI model to score loan applications. Analysis shows it approves a protected group at a substantially lower rate, even though applicants' protected attributes were never used as model inputs. Under nondiscrimination law, why can this still create legal exposure?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Liability is impossible here because excluding the protected attributes from the model inputs guarantees the model is legally compliant.Plausible but wrong: proxy variables can reproduce disparities, so omitting protected attributes does not guarantee compliance.
- Liability attaches only where a human underwriter has consciously intended to disadvantage the protected applicant group.Plausible but wrong: disparate-impact theory does not require proof of discriminatory intent.
- Nondiscrimination law simply does not reach automated lending decisions because the statutes predate machine-learning systems.Wrong: existing statutes apply to decisions however they are made, including by AI systems.
- Disparate-impact liability can arise from a facially neutral model that produces significantly unequal outcomes across groups. ✓Correct: nondiscrimination law reaches neutral practices that cause disproportionate adverse effects, regardless of intent.
The trap
Believing that removing protected attributes from inputs makes a model automatically nondiscriminatory. How to remember it
Neutral models can still violate nondiscrimination law through disparate impact when they produce significantly unequal outcomes.
How many of these would you get right?
One of 1581 AIGP questions on Certsqill. Take a free five-minute check and see your score per domain — not one number, but which section to open tonight.
Test your AIGP readiness — freeMore Understanding How Laws, Standards and Frameworks Apply to AI questions
- Reproducing copyrighted works to assemble a training set: Which intellectual-property consideration most →
- The prohibition on deceptive acts applies because: Under consumer-protection law, which theory applies most →
- A design-defect claim: Under product-liability law, which theory most directly fits? →
- All 394 Understanding How Laws, Standards and Frameworks Apply to AI questions →
Part of the Certsqill AIGP question bank · Understanding How Laws, Standards and Frameworks Apply to AI ·
Every answer, right and wrong, comes with its own explanation.
