OpenAI Releases Open-Source Model to Strip Personal Data from Text

OpenAI has launched the Privacy Filter, an open-source model engineered to identify and redact personally identifiable information (PII) within text.

Designed for teams that need to sanitize large volumes of text before downstream processing—whether for training proprietary AI systems or sharing data with external partners—the tool operates efficiently on modest hardware. Weighing in at 1.5 billion parameters with just 50 million active per request, it runs entirely locally on laptops or directly within web browsers, requiring zero cloud connectivity.

The model scans for eight distinct categories of sensitive data: names, physical addresses, email addresses, phone numbers, URLs, dates, account numbers, and secrets such as passwords or API keys. Unlike generative chatbots, it does not produce new content. Instead, it performs a single-pass analysis of the input, tagging each segment according to its classification. Backed by a 128,000-token context window, the system can process lengthy documents seamlessly without requiring segmentation.

Users can fine-tune the detection sensitivity, opting for an aggressive masking approach (prioritizing high recall at the cost of more false positives) or a conservative mode (reducing false alarms but potentially missing some items). Organizations with proprietary datasets also have the flexibility to further fine-tune the model for specialized use cases.

The Privacy Filter is publicly available under the Apache 2.0 license on both GitHub and Hugging Face, fully permitting commercial deployment and integration.

Clear Limitations for Sensitive Use Cases

OpenAI explicitly clarifies that this tool does not offer legal guarantees regarding anonymization or regulatory compliance. It should be viewed as one component within a broader data governance framework. The company also acknowledges several technical constraints: regionally uncommon or rare names are more prone to being missed, high-profile individuals or major organizations may occasionally be incorrectly flagged, and performance noticeably degrades when processing non-English or non-Latin script content.

For highly regulated sectors such as healthcare, legal services, finance, and human resources, OpenAI strongly recommends maintaining human-in-the-loop review. Additionally, detection categories cannot be modified dynamically at runtime, meaning teams requiring customized redaction policies will need to train their own fine-tuned versions.