What is the difference between PII redaction, masking, and tokenization for LLMs?

Question

Accepted Answer

Redaction irreversibly removes values ([REDACTED]). Masking partially hides them for humans. Deterministic tokenization replaces each unique value with a stable label like [EMAIL_1] so agents and support traces can still refer to the same entity — NoeticGuard’s default for LLM egress. Key takeaways: Use redaction when you never need the original in that pipeline. Partial masks are weak for LLM reasoning continuity. Vault-stable tokens keep debugging possible without sending raw PII to the model — and reuse the same placeholder when a value repeats.

Approach	What the model sees	Best when
Redaction	`[REDACTED]` / empty — irreversible	You never need the original in that pipeline
Masking	Partial reveal (`j***@…`) or placeholders	Human UIs; weak for LLM reasoning continuity
Deterministic tokenization	Stable `[EMAIL_1]` per vault session	Agents, tools, and support traces that must correlate entities

PII redaction vs masking vs tokenization for LLM egress