Representativeness bias occurs when a training dataset systematically over- or under-samples subpopulations relative to the population the deployed system will actually encounter — for example, a facial recognition dataset skewed toward lighter-skinned faces, or a multimodal medical dataset drawn predominantly from one demographic group. Because models learn statistical patterns from their training distribution, an unrepresentative dataset produces a model whose accuracy, calibration, and fairness properties degrade for underrepresented groups, even when aggregate accuracy metrics look acceptable.

This is precisely why aggregate accuracy is an insufficient safeguard: option B's framing — that bias doesn't matter "as long as predictions are accurate" — conflates overall accuracy with subgroup accuracy, and a model can post strong aggregate numbers while systematically failing specific populations. Option D is factually false; AI systems have no inherent neutrality — they inherit and can amplify whatever patterns (including societal biases) exist in their training data and objective function. Option C is also incorrect: mitigating representativeness bias is significantly cheaper and more effective when addressed at the data-collection and curation stage — through stratified sampling, bias audits, and diverse data sourcing — than after deployment, when it becomes a retraining and remediation problem, and by then real-world harm may have already occurred.