OpenAI releases open-source model that strips personal data from text

OpenAI has released Privacy Filter, an open-source model designed to detect and redact personal data in text.

According to OpenAI, Privacy Filter is built for teams that need to clean large volumes of text before processing it further, whether for training their own AI models or sharing data with third parties. The model is relatively small at 1.5 billion parameters, uses only 50 million active parameters per request, and runs on a laptop or even directly in a browser, OpenAI says. Running it on local hardware without any cloud connection is explicitly supported.

The model detects eight categories of sensitive content: names, addresses, email addresses, phone numbers, URLs, dates, account numbers, and other secrets like passwords or API keys. Unlike traditional chatbots, it doesn't generate new text. Instead, it makes a single pass through the input and labels which parts belong to which category. A 128,000-token context window lets it process long documents without splitting them up, according to OpenAI.

Users can adjust settings to control whether the model redacts aggressively (high recall, more false positives) or conservatively (fewer false positives, but more missed items). Teams with their own datasets can also fine-tune the model further.

Privacy Filter is available under the Apache 2.0 license on GitHub and Hugging Face, and commercial use is permitted.

Clear limits for sensitive use cases

OpenAI is clear that Privacy Filter does not provide any legal guarantee of anonymization or compliance. The model is meant to be just one layer in a broader data protection strategy. OpenAI itself lists several weaknesses: rare or regionally uncommon names are more likely to be missed, well-known public figures or organizations sometimes get incorrectly redacted, and performance drops with non-English text or non-Latin scripts.

For sensitive fields like healthcare, law, finance, or human resources, OpenAI explicitly recommends keeping human review in the loop. The label categories also can't be changed at runtime, meaning that teams that need a different policy will have to fine-tune the model.