OpenAI Releases Privacy Filter Model for Detecting and Masking Personal Data

OpenAI announced in a press release the release of Privacy Filter, an open-weight model designed to detect and mask personally identifiable information in text. The model aims to support developers in implementing stronger privacy safeguards in their AI workflows.

Privacy Filter is a compact model capable of context-aware detection of private data in unstructured text. It can run locally, allowing users to mask or redact sensitive information without sending data to external servers. The model supports up to 128,000 tokens of context and labels text spans across eight privacy categories, including personal names, addresses, emails, phone numbers, URLs, dates, account numbers, and secrets.

The model uses a bidirectional token classification architecture with a constrained decoding process, labeling all tokens in a single pass for faster performance. It achieved an F1 score of 96 percent on the PII-Masking-300k benchmark and 97.43 percent on a corrected version of the dataset. Developers can fine tune the model for domain specific use cases and customize masking behavior.

Privacy Filter is available under the Apache 2.0 license on Hugging Face and GitHub, with documentation covering architecture, taxonomy, evaluation, and limitations. OpenAI uses a fine tuned version of the model internally for privacy preserving workflows and encourages external feedback for further refinement.