OpenAI Launches Privacy Filter for PII Detection and Redaction

Caroline Bishop Apr 24, 2026 15:27

OpenAI unveils Privacy Filter, a model for detecting and redacting PII with state-of-the-art accuracy. Available now for developers.

OpenAI has introduced the Privacy Filter, a new open-weight AI model designed to detect and redact personally identifiable information (PII) in text. Released on April 23, 2026, the tool offers state-of-the-art accuracy for privacy workflows, setting a new benchmark for safeguarding sensitive data.

The Privacy Filter aims to address longstanding challenges in PII detection by moving beyond traditional rule-based systems. Unlike older tools that rely on deterministic patterns for formats like phone numbers or email addresses, this model uses advanced language understanding to analyze unstructured text and make context-aware decisions. For instance, it can differentiate between public and private information, a crucial capability in nuanced privacy scenarios.

One of its standout features is the ability to operate locally, meaning sensitive data can be processed directly on a user’s machine without being sent to external servers. This design reduces the risk of data exposure and is particularly valuable for privacy-sensitive applications in industries like healthcare, finance, and legal services.

The model is optimized for production use, with efficiency at its core. It can process up to 128,000 tokens of context in a single pass while maintaining high precision and recall rates. On the widely used PII-Masking-300k benchmark, the Privacy Filter achieved an F1 score of 96%, climbing to 97.43% on a corrected version of the dataset. These scores highlight its ability to catch subtle and complex privacy breaches while minimizing false positives.

OpenAI has made the model available under an Apache 2.0 license via platforms like Hugging Face and GitHub, encouraging developers to experiment, fine-tune, and deploy it for their specific needs. The model supports eight PII categories, including private_person, private_address, private_email, and account_number. It’s configurable for various workflows, allowing users to balance precision and recall based on their operational priorities.

Despite its capabilities, OpenAI acknowledges that the Privacy Filter isn’t a standalone solution for regulatory compliance or policy review. It’s a tool designed to complement broader privacy-by-design strategies. The model may require further domain-specific evaluation or fine-tuning to adapt to different industries, languages, and naming conventions.

This release is part of OpenAI’s broader push to enhance AI-driven privacy protections. The company has emphasized that the goal is for AI systems to learn from data without compromising individual privacy. By open-sourcing the Privacy Filter, OpenAI aims to empower developers and organizations to integrate robust privacy safeguards into their workflows.

The Privacy Filter represents a significant step forward in privacy technology, especially as the AI industry faces increasing scrutiny around data security and ethical practices. For developers and businesses looking to strengthen their PII-handling capabilities, the model’s availability marks a crucial opportunity to adopt cutting-edge tools that prioritize user privacy.

• ai
• privacy
• openai
• pii
• data security