A manager wants to ensure that only quality data is used in their AI model.

Training a facial-recognition system predominantly on one demographic creates representation bias. The model receives insufficient examples from other groups and will probably perform less accurately for those populations, producing systematically unequal outcomes. This is directly associated with fairness because model performance varies according to demographic characteristics. The other scenarios describe serious data-quality problems, but their primary effects differ. Missing merchant details can reduce fraud-detection accuracy, incorrectly encoded text introduces corruption, and bot traffic distorts customer-behavior signals. Those defects may degrade overall performance without necessarily disadvantaging a protected or underrepresented group. Responsible AI development requires representative datasets, subgroup-level evaluation, documented data provenance, and ongoing monitoring for unequal error rates. Therefore, the facial-recognition dataset presents the clearest and most direct risk of an unfair and biased outcome.