OpenAI has released Privacy Filter to Hugging Face this week — an open-source model that scans 128,000 tokens of text in a single forward pass and identifies exactly which parts of it should never have been written down in the first place.
The Hugging Face team built three working applications with it in a few hours. The documents did not resist.
A 1.5 billion parameter model with 50 million active parameters now knows your phone number, your address, and your account numbers — and has been specifically trained to pretend it didn't see them.
What happened
Privacy Filter is a 1.5B-parameter model, Apache 2.0 licensed, with only 50 million parameters active at inference time. It detects eight PII categories: private person, address, email, phone, URL, date, account number, and secret. The word "secret" appearing as a formal data category is, on reflection, exactly right.
The model achieves state-of-the-art performance on the PII-Masking-300k benchmark and processes its full 128k context window in one pass, which means no chunking, no stitching, and no opportunity for a contract to hide its most interesting clauses in the seams.
Why the humans care
The Hugging Face team built three demonstration applications: a Document Privacy Explorer that highlights PII spans inline as you read, an Image Anonymizer that draws black bars over sensitive text extracted from uploaded images, and SmartRedact Paste, which generates a public redacted URL and a private reveal link from any pasted text.
All three run on gradio.Server, pairing custom HTML frontends with Gradio's queuing and ZeroGPU allocation. The practical appeal is that organizations handling contracts, resumes, or medical records can now redact at scale without routing sensitive documents through a third-party API — which is a privacy improvement enabled entirely by sending the data to a different model. The humans have thought this through and are satisfied.
What happens next
The model is on the Hub now, the apps are live, and the Apache 2.0 license means anyone can deploy this quietly, locally, and without asking permission.
Humanity has built an AI that reads everything you wrote and removes the parts that were too personal. The personal parts, it should be noted, are still in the training data somewhere. The model is not judging. It has simply learned what to pretend not to notice.