OpenAI Launches Privacy Filter: AI for Personal Data

OpenAI has recently introduced "Privacy Filter," an open-source artificial intelligence model designed to anonymize personal data in texts. This model, which operates locally, is capable of automatically detecting and masking sensitive information before any further processing.

Features and Capabilities

Privacy Filter can recognize eight categories of personal data, including names, addresses, email addresses, phone numbers, URLs, dates, account numbers, and other secrets like passwords or API keys. Unlike traditional chatbots, it does not generate new text. Instead, it performs a single pass through the input and labels the parts according to their category.

The model is designed to handle long documents with a context window of 128,000 tokens, allowing it to process texts without splitting them. The model is relatively compact, with 1.5 billion parameters, but only uses 50 million active parameters per request. This enables it to run on laptops or directly in a browser, without requiring a cloud connection.

Usage and Adjustments

OpenAI designed Privacy Filter for teams that need to clean large volumes of text before further processing, whether for training their own AI models or sharing data with third parties. Privacy Filter is available under the Apache 2.0 license on GitHub and Hugging Face, allowing for commercial use. Users can adjust the model's sensitivity for more or less aggressive redaction, depending on their specific needs. Teams with their own datasets can also fine-tune the model to better meet their requirements.

Limitations and Recommendations

Despite its capabilities, OpenAI emphasizes that Privacy Filter does not guarantee compliance with anonymization laws. The model has weaknesses, particularly in detecting rare names or public figures, and its performance declines with non-English texts or non-Latin scripts. The model is intended to be one layer within a broader data protection strategy.

For sensitive fields such as healthcare, law, finance, or human resources, OpenAI recommends maintaining human oversight in the process to ensure adequate data protection. The label categories cannot be modified in real-time, requiring model adjustments for different policies.