A university lecturer teaches about "data minimization in AI feature design." Which of the following is the best application of this principle when building a loan-underwriting AI?

• A.Minimize the number of model parameters to reduce inference computation costs; privacy regulators may not compel model deletion: the FTC's algorithmic disgorgement authority was struck down in 2023 and no equivalent power exists in the EU, meaning a model trained on unlawful data remains a lawful asset
• B.Collect only features demonstrably necessary for credit risk prediction (e.g., repayment history, debt-to-income ratio) and exclude features that are not predictively necessary (e.g., social media activity, shopping preferences)
• C.Use all available customer data and let the model learn which features are predictively useful during training
• D.Minimize data storage costs by compressing all training data with lossless compression; opt-out signals are enforced at the tokenizer level, meaning a model cannot ingest text carrying a do-not-train tag even if the pipeline tries