OpenAI Privacy Filter

OpenAI Privacy Filter is a text redaction model for detecting personally identifiable information and secrets. You give it text, and it marks the spans that look private so they can be masked, reviewed, or handled separately.

The useful angle is where it fits in a workflow. You can run a customer support transcript, internal note, chat export, log file, or document batch through the model before that text is shared with another tool. Instead of treating privacy cleanup as a manual pass, you get a local first filter that can catch names, emails, phone numbers, account numbers, dates, URLs, and secret-like strings.

It is also built for longer inputs than a typical small redaction widget. The model card describes a 128,000-token context window, browser and Python examples, and runtime controls for choosing a stricter or looser masking behavior. That makes it more interesting for document cleanup and data preparation than a one-off named-entity demo.

Start with the Hugging Face Space and paste a short but realistic sample. Use something messy: a support message with a name, email, phone number, account number, API-key-like string, date, public company name, and a few normal sentences around them. The first thing to check is not only what it catches, but whether it avoids hiding useful context that is not actually private.

If the browser result is useful, move to the model page for Transformers or Transformers.js examples. For a more reproducible local path, use the OpenAI GitHub repo and its `opf` command-line tool. That path supports one-shot redaction, file input, pipes, evaluation on labeled data, and finetuning when your privacy policy does not match the default labels.

Do not treat this as an anonymization or compliance guarantee. It can miss uncommon identifiers, over-redact public entities or harmless strings, and behave differently on non-English text or domain-specific formats. Use it as one layer in a privacy workflow, then test it on the kind of text you actually handle.

• Redact customer messages before sending them into an analytics or AI workflow.
• Clean internal notes, transcripts, support logs, or documents before sharing them with a vendor.
• Build a first-pass privacy filter for retrieval, search indexing, or dataset preparation.
• Compare strict and loose masking behavior on your own examples before deciding where human review is needed.
• Fine-tune the workflow when your organization has a different definition of what should be masked.

Try the demo

View model page