Introducing OpenAI Privacy Filter

OpenAI today announced the release of the Privacy Filter, an open‑weight model aimed at detecting and redacting personally identifiable information (PII) in text. The model, which contains 1.5 billion total parameters and 50 million active parameters, is designed for high‑throughput privacy workflows and can run locally, allowing PII to be masked without sending data to a server. Built as a bidirectional token‑classifier with span decoding, it processes inputs in a single forward pass and supports up to 128,000 tokens of context. OpenAI reports that the released version achieves a 96% F1 score on the PII‑Masking‑300k benchmark and 97.43% on a corrected version that accounts for annotation issues. The model is fine‑tuned from an autoregressive checkpoint and trained on a mix of public and synthetic data to capture realistic and challenging privacy patterns. OpenAI notes that while the model offers frontier‑level performance, it is not a substitute for policy review or compliance certification, and that fine‑tuning may be necessary for domain‑specific accuracy. The Privacy Filter is available under the Apache 2.0 license on Hugging Face and GitHub, with accompanying documentation on architecture, taxonomy, and limitations to aid developers in integrating the tool into privacy‑by‑design systems. The company emphasizes that the model is intended for experimentation, customization, and commercial deployment, and that it can be adapted to different data distributions and privacy policies. This release reflects OpenAI’s broader effort to provide practical infrastructure for building AI safely and to support stronger privacy protections across the ecosystem.