A county wildlife-rehab desk wants a helper on visit notes but must not keep raw phone numbers in the training or retrieval pile. Which handling belongs in the data path?
Select an answer to reveal the explanation.
Short Explanation
Think of visit notes that must not keep raw phone numbers in the training or retrieval pile. Minimize, redact, or tokenize first. Apologizing later is not a data path.
Full Explanation
The county wildlife-rehab desk wants a helper on visit notes but must not keep raw phone numbers in the training or retrieval pile. Minimization, redaction, or tokenization belongs in the data path before notes become AI data. Removing or substituting direct identifiers reduces privacy risk in embeddings, logs, and model memorization without blocking useful clinical or operational text the desk needs.
Keeping raw numbers in the pile and apologizing later fails because apology is not a preventive control and deletion from vectors is harder after the fact. Amazon QuickSight-style charts fail because dashboards do not tokenize phone numbers in training corpora. Raising temperature so phone numbers drop out of the notes fails because sampling settings do not scrub stored data or retrieval chunks.
Exam caveat: tokenization must be consistent if rejoining to operational systems is required, and keys or mapping tables need the same protection as the original numbers under least privilege. Operational check: insert a pre-ingest scrubber that detects phone patterns, redacts or tokenizes them, sample the AI corpus for residual raw numbers, and block training or indexing jobs when the scrubber reports failures above threshold.