OpenAI Releases Privacy Filter AI Model for Personal Data Detection and Redaction

OpenAI has launched Privacy Filter, an open-weight model designed to detect and mask personally identifiable information in text. The model achieves state-of-the-art performance with a 97.43% F1 score on the corrected PII-Masking-300k benchmark. With 1.5 billion total parameters and 50 million active parameters, Privacy Filter supports context lengths up to 128,000 tokens and operates locally to keep sensitive data on-device. It identifies eight categories of private information including account numbers, secrets, personal identifiers, contact details, addresses, and private dates. The bidirectional token-classification model uses span decoding with BIOES tags for coherent masking boundaries. Available under Apache 2.0 license on Hugging Face and GitHub, developers can fine-tune the model for specific use cases in training, indexing, logging, and review pipelines. OpenAI developed Privacy Filter through a multi-stage process involving privacy taxonomy definition, model conversion from pretrained checkpoints, and training on mixed public and synthetic data. The release aims to strengthen privacy protections across AI ecosystems while acknowledging limitations in high-sensitivity domains requiring human review.