OpenAI releases Privacy Filter, an open model for stripping personal data from text

OpenAI has published Privacy Filter, an open model built on a distilled decoder that finds and removes personally identifiable information (PII) from text. Although its weights total 1.5 billion parameters, only 50 million are active during inference, which lets it run directly in the browser with no server infrastructure.

What Privacy Filter is

Privacy Filter is a specialized language model built for one specific job: automatically detecting and redacting personally identifiable information (PII) in text. It is not a general-purpose chat assistant but a utility tool, focused solely on finding sensitive information and replacing it with standardized placeholders.

The specialization is an advantage: a narrowly tuned model handles the task better than general-purpose LLMs, which often miss unusually phrased personal data or make mistakes in complex contexts.

The model is based on a distilled decoder: a large teacher model passes its knowledge to a compact student model through distillation. The result is high detection accuracy with significantly lower compute requirements. The model is openly available, so any company can build it into its own pipelines without sending data to OpenAI's servers.

Architecture: 50 million out of 1.5 billion

The key technical detail is the gap between the total parameter count (1.5 billion) and the parameters actually activated for each token (50 million). This approach is typical of sparse-activation architectures: different neural blocks specialize in different aspects of the task and switch on selectively, depending on the input.

That makes Privacy Filter practical in resource-constrained scenarios:

• Browser: compatible with WebAssembly and ONNX, so data never leaves the user's device
• Edge devices: 50M active parameters allow it to run without a GPU on laptops and smartphones
• Self-hosted: the model fits entirely inside a company's own infrastructure
• CI/CD pipelines: fast inference with no cloud dependencies and no extra costs

What it can detect

Privacy Filter recognizes a broad set of personal data categories, covering key requirements of GDPR, LGPD and CCPA:

• Names, surnames, initials (including contextual recognition without explicit markers)
• Addresses, postal codes, geographic coordinates
• Phone numbers and email addresses
• Identity documents: passports, SSNs, INN (Russian tax ID numbers), driver's licenses
• Financial data: card and bank account numbers
• Medical identifiers

The model does not just flag PII fragments; it also replaces them with standard placeholders: [NAME], [ADDRESS], [PHONE]. The output text is immediately ready for further processing with no manual post-processing.

Regulatory context

Regulatory pressure on personal data is growing worldwide. GDPR in Europe, LGPD in Brazil and CCPA in California all require companies to handle sensitive information carefully. Most commercial automatic anonymization solutions either fell short on quality or required sending data to the cloud, which itself undermined the point of confidentiality.

Privacy Filter fills that gap: an open model with browser compatibility that a small team can build into a product in a day without sacrificing user privacy.

What this means

OpenAI keeps investing in open infrastructure alongside its commercial flagships. Privacy Filter shows that the company sees a market not only in API access to GPT but also in utility tools that address specific operational needs. It signals that enterprise-grade open-source tools for AI data security are becoming the norm.

For businesses, it is a ready-made solution to the anonymization problem, with no need to build from scratch and no dependence on the cloud.