Introducing OpenAI Privacy Filter

Summary

OpenAI has released Privacy Filter, an open-weight model designed for context-aware detection and redaction of personally identifiable information in unstructured text. The 1.5B-parameter model, with 50M active parameters, operates as a bidirectional token classifier that labels sequences in a single pass before decoding coherent spans via a constrained Viterbi procedure. It supports eight privacy categories, including account numbers, secrets such as passwords and API keys, personal names, dates, emails, and phone numbers, and distinguishes public from private references based on surrounding context rather than rigid pattern matching.

The model runs locally and processes long inputs efficiently, enabling redaction without transmitting unfiltered data to external servers. This architecture suits high-throughput pipelines in training data preparation, logging, indexing, and review workflows. OpenAI reports that a fine-tuned version of the model is already used internally, and the released checkpoint reaches 97.43 percent F1 on a corrected version of the PII-Masking-300k benchmark, with 96.79 percent precision and 98.08 percent recall.

Privacy Filter is distributed under the Apache 2.0 license on Hugging Face and GitHub, allowing fine-tuning on domain-specific data. The company notes that performance can vary across languages, naming conventions, and contexts not well represented in training, and that the tool is not a substitute for policy review or human oversight in high-stakes domains.

Why it matters

Directly actionable for Dutch/EU security and privacy teams needing GDPR-compliant PII redaction that runs on-premise without sending data externally. Offers fine-tunable, context-aware detection superior to rule-based tools, addressing key privacy risks in AI systems.