OpenAI Privacy Filter Model Card

OpenAI releases the Privacy Filter model card, introducing a compact bidirectional token classification model for detecting and redacting personally identifiable information (PII) and confidential information in unstructured text. The model is converted from an autoregressive pretraining checkpoint into a banded-attention bidirectional classifier; a single forward pass labels the input sequence, and then a constrained Viterbi decoder generates coherent redaction spans across eight privacy categories and provides configurable precision-recall trade-off points. The model has 1.5B total parameters, 50M activated parameters per token, and a 128,000-token context window, and is designed for efficient local deployment and domain-specific fine-tuning. The model card states that its positioning is as a configurable data minimization component in a layered privacy workflow, rather than an anonymization or compliance guarantee.