Subtitle: when a free tool is genuinely enough — and where the one-shot model breaks
If you just need to strip emails from one CSV, once, a free scrubber is the right answer. This post is about what happens after "once" — and it starts by conceding the biggest point: detection itself is now free.
Detection became a commodity
In April 2026, OpenAI released Privacy Filter, an Apache-2.0 PII masking model that runs in a browser or on a laptop. It covers eight PII categories, and it's genuinely good. OpenAI's own framing is worth quoting: it is a "redaction aid, not a safety guarantee" — a caveat every honest tool in this category should echo, ours included.
So if free models and free scrubbers can find PII, what's left to pay for? The answer isn't better detection. It's everything a stateless, one-shot tool structurally cannot do.
What one-shot tools can't do
- Consistency. A scrubber that replaces every name with
[REDACTED]destroys the very thing you wanted to analyze: "customer #3 complained four times" becomes "someone complained, four separate times, maybe." Useful anonymization maps the same person to the same alias — PERSON_001, every row, every time. - Reversibility. You send the safe copy to ChatGPT and it answers: "PERSON_001 needs a follow-up." Now what? Without a stored mapping, that answer is decoration. The round-trip — alias out, real name back in — is where anonymized data becomes usable work.
- Memory. Monday's file and Friday's file should agree on who PERSON_001 is. One-shot tools can't, because they keep no state. (State is also a responsibility: in our view it belongs encrypted, on the user's machine — not on a vendor's server.)
- Structure. Analysis needs row order, column order, empty cells, and numeric columns intact. Tools built for redacting prose tend to flatten tables; tools built for tables should hand back a file your pivot table still understands.
The upload paradox
One more question to ask any scrubber, free or paid: where does the processing happen? If the answer is "our server," you have uploaded the very file you were trying to protect — before any protection was applied. That's not hypothetical risk; it's a second copy of your original data on someone else's infrastructure.
The test is simple and takes two minutes: open your browser's DevTools network tab (or go offline) and process a file. Requests carrying your file content, or a tool that stops working offline, tell you everything.
An honest buying guide
- One file, one time, no free-text columns → use a free scrubber or a find-and-replace. Genuinely fine.
- Recurring files, same shape → you need consistency; check whether your tool remembers mappings, or you'll rebuild them weekly by hand.
- AI analysis whose output returns to real work → you need reversibility. This is the workflow free tools structurally skip.
- Data that must not leave your machine → local processing only, and verify the claim instead of trusting it.
We build Data Alias around exactly this gap — in-browser processing (verifiable in two minutes), consistent aliases, an encrypted local vault for cross-file memory, and de-aliasing for AI responses. And in the spirit of the honesty this category owes you: no automatic detector catches every sensitive value, ours included — which is why every export passes through a human review step.
The paid layer in this category was never detection. It's the workflow that lets protected data keep doing work.