OpenAI privacy tool

OpenAI said on April 22 it is open-sourcing Privacy Filter, a model that detects and redacts personal information in text before that data is reused elsewhere. (openai.com) In plain terms, the system reads text like a highlighter looking for names, phone numbers, account numbers, addresses, and other sensitive details, then masks those spans automatically. OpenAI describes it as a bidirectional token-classification model built for personally identifiable information detection and redaction. (openai.com) OpenAI released the weights under an Apache 2.0 license on GitHub and Hugging Face, which allows commercial use and modification. The company said the model has 1.5 billion parameters in total, with 50 million active parameters, and can run in a web browser or on a laptop. (github.com, huggingface.co) The model does not generate prose like ChatGPT. It labels each token in an input sequence, then uses a constrained decoding step to turn those labels into coherent redacted spans across documents as long as 128,000 tokens. (github.com, cdn.openai.com) OpenAI said the release is aimed at developers and companies that want to scrub sensitive data before sending prompts into chatbots, application programming interfaces, or internal workflows. The company’s model card says users can tune the system for higher precision or higher recall depending on whether they want fewer false alarms or fewer missed secrets. (openai.com, cdn.openai.com) That fits a broader push by OpenAI to show privacy controls earlier in the artificial intelligence pipeline, not only at the chatbot stage. In a separate privacy paper, OpenAI said it already uses internal filtering to mask personal information that may appear in training data. (cdn.openai.com, openai.com) The company is also making a strategic point about where the tool runs. VentureBeat reported that Privacy Filter is designed for on-device use, which lets organizations inspect and deploy the model inside their own environments instead of sending raw files to a third party first. (venturebeat.com, github.com) OpenAI’s own materials describe the release as text-focused, even as some outside reports framed it as covering both text and image prompts. The official announcement, GitHub repository, and model card all define Privacy Filter as a text redaction model rather than a vision system. (openai.com, github.com, cdn.openai.com) For OpenAI, the release adds one more open-weight tool alongside its newer open-model push. For developers handling resumes, support logs, medical notes, or financial records, it offers a way to blank out the obvious identifiers before a chatbot ever sees the text. (openai.com, openai.com)