OpenAI Privacy Filter

Privacy filters are the software equivalent of a black marker: they scan text for names, phone numbers, emails and other identifying details before anything is sent onward. OpenAI said on April 22 it open-sourced a model called Privacy Filter that does that masking locally, before a downstream chatbot sees the raw text. (openai.com) OpenAI described Privacy Filter as an open-weight model for detecting and redacting personally identifiable information in text, and said it is already using a fine-tuned version in its own privacy-preserving workflows. The company released code and model assets on GitHub under an Apache 2.0 license. (openai.com, github.com) The model does not write an answer the way a chatbot does. OpenAI said it labels each token in an input in a single pass, then groups those labels into spans to decide what should be masked. (github.com, openai.com) That design targets a practical problem for companies using large language models: the riskiest data often appears before the model ever starts answering. OpenAI said Privacy Filter is meant for training, indexing, logging and review pipelines where teams want to strip sensitive data before it leaves their own machines. (openai.com, github.com) OpenAI said the model is small enough to run in a web browser or on a laptop, with 1.5 billion total parameters and 50 million active parameters at a time. The repository also lists a 128,000-token context window, which lets it process long documents without splitting them into many smaller chunks. (github.com) The company said the model is built to do more than pattern matching for obvious formats like email addresses and phone numbers. In its release note, OpenAI said the system uses context to separate information about public figures from information tied to private individuals. (openai.com) OpenAI said the released version reaches state-of-the-art performance on its PII-Masking-300k benchmark after correcting annotation issues it found during evaluation. The model card also lists multilingual and adversarial testing, alongside a section warning against over-reliance and cautioning about high-risk deployments. (openai.com, cdn.openai.com) The release also shifts some responsibility back to developers. OpenAI said teams can run, inspect and fine-tune the model in their own environments, which makes privacy scrubbing a deployment choice builders have to configure before prompts reach a larger model. (openai.com, github.com) In practice, the pitch is simple: redact first, ask questions second. OpenAI’s new filter turns that step into a small model that can sit in front of a chatbot instead of behind it. (openai.com, decrypt.co)