OpenAI has launched Privacy Filter, a new tool designed to scrub sensitive personal information from text before it's processed by AI models like ChatGPT. Released under the Apache 2.0 license, this open-source model can be downloaded, modified, and used freely by anyone.
The 1.5 billion-parameter model runs locally on a user's computer, ensuring that private data never leaves the device. It scans for eight categories of personal information, including names, addresses, emails, phone numbers, URLs, dates, account numbers, and secrets like passwords and API keys.
Privacy Filter replaces identified sensitive data with generic placeholders such as [PRIVATE_PERSON] or [ACCOUNT_NUMBER]. Unlike simpler pattern-matching tools, it analyzes the surrounding text for context, allowing it to distinguish between personal names and brands, or between addresses of individuals and businesses.
OpenAI reports the model achieves 96% F1 score on the standard PII-Masking-300k benchmark out-of-the-box, with a corrected version reaching 97.43%. This local processing capability is crucial for users handling sensitive documents like tax returns, medical records, or client emails, safeguarding data privacy.
While effective, OpenAI cautions that Privacy Filter is not a complete anonymization tool and should not be the sole defense for highly sensitive environments. It can miss unusual identifiers and performs unevenly across languages, emphasizing that 96% accuracy is not 100%.