AI Bias & Discrimination: Complete Management Guide | Quiz

Why: Two or more Yes answers to the four risk questions (affects humans and uses demographic proxies and trained on historical decisions and high-risk under EU AI Act) means mandatory bias management is required. Context: This threshold helps organizations quickly identify which AI systems need formal bias programs. Hiring AI typically answers Yes to multiple questions making it high-risk. Remember: Two or more Yes answers equals mandatory bias management program.

Why: Zip code correlates strongly with race due to historical housing discrimination (redlining) – using it as a feature allows the model to indirectly discriminate based on race without ever seeing race as an input. Context: This is algorithm amplification through proxy variables. The model is not given race but finds zip code as a proxy that correlates with it. This still creates disparate impact liability. Remember: Zip code is a proxy for race due to housing discrimination history.

Why: Counterfactual testing changes only the protected attribute (like changing a name from John to Jamal) while keeping everything else identical – if the outcome changes you have found potential bias. Context: This technique reveals discrimination that aggregate metrics might miss. It directly tests whether the model treats people differently based on protected characteristics. Remember: Change only the protected attribute – if outcomes change you found bias.

Why: Mathematical proofs show you cannot simultaneously achieve all fairness metrics because different definitions of fairness mathematically conflict with each other. Context: You must choose which fairness definition matters most for your specific use case and stakeholder values. Demographic parity and equal opportunity and equalized odds cannot all be satisfied at once. Remember: All fairness metrics cannot be achieved simultaneously – you must choose.

Why: A feedback loop occurs when biased AI decisions create data that reinforces the original bias – predictions become self-fulfilling prophecies perpetuating the disparities they measure. Context: The predictive policing example shows this clearly. More police in minority neighborhoods means more arrests there which generates data confirming those areas are high-crime which justifies more policing. Remember: Biased decisions create biased data which creates more biased decisions.

Why: Models are excellent at finding proxy variables that correlate with protected attributes – names and schools attended and zip codes and even writing style can serve as proxies. Context: The classic example is zip code predicting credit default risk. Zip code correlates with race due to historical housing discrimination so the model indirectly discriminates without ever seeing race as an input. Remember: Remove the attribute but models find proxies – correlation survives removal.

Why: Errors are random and symmetric affecting all groups equally while bias is asymmetric and directional consistently disadvantaging the same groups. Context: If an AI makes 5% false positives across all groups that is error. If it makes 20% false positives for Group A and 5% for Group B that is bias. This distinction determines legal liability. Remember: Error is random and equal – bias is directional and unequal.