OpenAI Erases PII

OpenAI has released a Privacy Filter model that detects and removes personally identifiable information (PII).

The model is designed to detect names, dates, account numbers, credit card numbers, email addresses, passwords, and more, and users can also fine-tune it to fit their own policies.

The model can run locally, so the original data never leaves the environment, and OpenAI stated that it internally uses a fine-tuned version of this model for data minimization purposes as well.

However, OpenAI made clear that this tool is not an anonymization tool, not a compliance certification, and not a substitute for policy review in high-risk environments. It explained that in heavily regulated fields such as law, healthcare, and finance, human review is still required.

The model was trained on public and synthetic data, and OpenAI said it will further refine the detection scope and usage based on developer feedback. At the same time, experts have pointed out that large models can reconstruct sensitive information through inference even after explicit identifiers are removed, suggesting that PII filtering alone may not be sufficient.