OpenAI open-sources privacy filter
- OpenAI published a small model that masks sensitive information before content is sent to ChatGPT. (decrypt.co) - It also launched ChatGPT for Clinicians with promises not to use conversations for training and optional HIPAA compliance for eligible users. (firstpost.com) - These moves signal that privacy tooling and contractual safeguards are now product features, not just legal afterthoughts. (decrypt.co)
OpenAI released Privacy Filter — an open‑weight model to mask personally identifiable information before text is sent to ChatGPT, and it also unveiled ChatGPT for Clinicians with explicit data‑use rules. (openai.com) The company published Privacy Filter on April 22, 2026 under an Apache‑2.0 license and pushed the repository and tools to GitHub and Hugging Face. (openai.com) Privacy Filter is a bidirectional token‑classification model built from the gpt‑oss family; OpenAI says it’s 1.5 billion parameters in total with about 50 million active parameters and a 128,000‑token context window, and it can run locally on laptops or in browsers. (github.com) OpenAI reported the model achieves state‑of‑the‑art results on the PII‑Masking‑300k benchmark; independent reporting of OpenAI’s evaluation lists F1 scores around 96% (rising to ~97.4% on a revised dataset). (openai.com) Separately, OpenAI launched ChatGPT for Clinicians and said the product is free for verified U.S. physicians, nurse practitioners, physician assistants, and pharmacists, offering documentation, medical research, clinical search, and reusable workflows. (openai.com) OpenAI and its product pages state conversations in the clinical space are not used to train its foundation models, and the company offers optional HIPAA support through Business Associate Agreements for eligible accounts. (help.openai.com) PII detection and redaction tooling already exists in the market — examples include Microsoft’s open‑source Presidio project and Amazon Comprehend’s managed PII detectors — but OpenAI’s release packages a small, tunable model and CLI for on‑device workflows under a permissive license. (github.com) OpenAI says it uses a fine‑tuned Privacy Filter internally and warns the model can err on uncommon or ambiguous identifiers, recommending human review in high‑sensitivity legal, medical, and financial workflows. (openai.com) Some commentary points out PII redaction is a longstanding problem with existing open and commercial tools; others flagged that making redaction and contractual safeguards available as product features changes how companies operationalize privacy protections. (hackernoon.com) The code and CLI are available now on OpenAI’s GitHub and Hugging Face, and verified U.S. clinicians can enroll in ChatGPT for Clinicians beginning with the April 22, 2026 rollout. (github.com)
