OpenAI Privacy Parser (github.com)

OpenAI has launched a new tool called Privacy Parser, designed to extract personally identifiable information (PII) from text, effectively enabling users to audit potentially sensitive data. This model operates on the same framework as its counterpart, Privacy Filter, which masks PII to prevent data leaks. While Privacy Filter focuses on protecting data, Privacy Parser serves the dual purpose of facilitating offensive strategies by allowing attackers to identify PII from compromised sources, such as data dumps or logs. This dual-use nature highlights the importance of responsible AI practices within the community. Privacy Parser utilizes a 1.5 billion parameter model that can identify eight categories of PII, including private emails, phone numbers, addresses, and account numbers, producing structured output rather than redacted text for enhanced clarity. It combines advanced machine learning techniques with a regex backstop to ensure accurate parsing, achieving a latency of around 600 milliseconds on CPU and a high F1 score of 0.929 in tests. The release, accompanied by user-friendly installation instructions, underlines OpenAI's commitment to transparency while also raising significant ethical considerations in the field of AI/ML regarding the potential misuse of such powerful tools.