There's Now a Free Tool That Strips PII Before Your AI Sees It. Here's When to Use It.
OpenAI released a free, on-device model that catches personally identifiable information before text reaches any server. 96% accuracy across 8 categories — and the 4% it misses is where your judgment still matters.
By Patin Team · Examples are illustrative composites
The main reason teams hesitate before pasting internal documents into AI tools is not that they don't trust the output. It is that the input contains names, email addresses, employee records, and contract details that were never meant to leave the building. There is now a free tool that screens for that information locally — before any text reaches a server — and strips identifying details before the content ever hits your AI tool.
What OpenAI released
On April 22, OpenAI published a free, open-weight model — 1.5 billion parameters — built specifically to detect and redact personally identifiable information from text before it leaves your device. The model runs entirely on-device, which means the content it screens never travels to an external server during the detection step. TLDR AI reported on the release the same day.
The model identifies eight categories of personally identifiable information: names, email addresses, phone numbers, dates of birth, financial account identifiers, and three others. Across all eight, it achieves a 96% F1 accuracy score — a standard measure that accounts for both catching real identifiers and avoiding false positives on words that merely resemble them.
What to do differently from Monday
The 96% accuracy figure means roughly four out of every hundred identifiable pieces of information will pass through undetected. For a document with 50 names and addresses, that is two items. For a large export with 1,000 identifiers, it is forty.
That rate is low enough to make the filter genuinely useful as a first pass: employee survey analysis, customer feedback synthesis, draft reports built from meeting notes. It catches most of what a manual skim misses on a busy afternoon. The 4% is not a reason to skip it — it is a reason to keep a spot-check step in place before the processed content goes anywhere external.
If your workflow involves legal filings, medical records, or documents with a compliance sign-off requirement, the filter still earns its place in the process — it speeds up the human review step rather than replacing it.
What this looks like in practice
A training coordinator at a 75-person logistics company runs monthly employee pulse surveys across three regional offices. The survey platform exports anonymous responses, but employees regularly mention colleagues and managers by name when describing specific situations. Before running any thematic analysis in an AI tool, she exports the data through the PII filter first. It removes the names she expected, and catches a direct email address embedded in one free-text comment — something she would likely have missed skimming 200 responses on deadline.
The filter does not change her workflow in any fundamental way. It adds four minutes. What changes is her confidence that the content she pastes into an AI tool is not carrying information that could identify individuals to the model.
What the filter does not catch
A legal operations manager at a 300-person consulting firm handles document reviews that include client references. Her concern is not just PII in the technical sense — names, account numbers, dates of birth — it is commercially sensitive context: deal valuations, unreleased product roadmaps, acquisition targets, internal project codes.
The PII filter does not flag any of that. It screens for identifiable personal information, not for confidential business information. Those are different categories with different risk profiles, and treating the filter as a general "sensitive information" screen is a mistake that looks reasonable until something slips through.
The practical check before any document enters an AI tool: run the filter for compliance exposure, then apply your own judgment to commercial sensitivity as a separate pass. The tool handles one of the two categories. Knowing which one is the skill.
The actual change this week
Teams now have a free, local option for the PII screening step — the kind of step that gets skipped on a busy day or handled with a fast manual search. Using it does not guarantee compliance. Not using it is a consistent way to miss the identifiers a skim does not catch.
Put this into practice
Reading is a start — but skill comes from doing. Try these drills now.
Reading about it only gets you so far
Patin turns this into five-minute drills that score what you write and tell you why. It's in closed beta — join the waitlist and we'll email you when your cohort opens.
Just want the writing? .
Keep reading on this
Your AI Vendor's Data Policy Keeps Changing. It's the Same Gap Five Times.
Five AI vendors made five different calls on your data this year — scraping, toggle resets, and a disconnect that didn't delete anything. One check catches all of it.
5 min readMeta's AI Terms Change October 30. ChatGPT Now Targets Ads From Your Prompts.
Meta's updated business terms, effective October 30, let it scrape content tied to any business account for AI training, with no opt-out — and the same week, ChatGPT widened ad targeting to use your prompts. Here's what to check before the deadline.
4 min readEvery AI Fix You've Already Made Has the Same Expiration Date
A privacy toggle that reset itself, a platform war over agent permissions, and a policy written after a UN warning all look like separate problems with separate fixes. They share one flaw: each was treated as something you set once, and none of them actually work that way.
5 min read