In neural network-based credit scoring, the model ' s complexity often obscures the relationship between specific inputs and final outcomes. If input validation and anomaly detection are weak, " biased inputs " or " noise " from various sources can seep into the training and inference pipelines. According to the ISACA AAIA™ framework, these biased inputs can cause the model to learn discriminatory patterns (e.g., correlations between ethnicity and creditworthiness) that are hidden within large datasets. This leads to unethical and potentially illegal credit decisions. While overfitting (Option B) and costs (Option C) are technical concerns, the systemic risk of automated discrimination via unchecked inputs poses the most significant threat to the organization ' s legal standing and reputation.