OpenAI launched this week an open-weight model called Privacy Filter, designed to strip personally identifiable information (PII) from text directly on a user’s device.
As TNS contributor Meredith Shubel reports, this 1.5-billion parameter model will allow developers to sanitize sensitive data locally on a laptop or browser, ensuring it never touches the cloud.
The launch of this model — interestingly, OpenAI uses a version of it in its own privacy-preserving workflows — serves as a tactical guardrail. It arrives alongside the company’s ChatGPT-5.5, which debuted Thursday.
But as OpenAI introduces what it calls a “new class of intelligence” (GPT 5.5), the larger concern remains: Can a local filter truly protect us from an engine designed to know everything?
OpenAI's new Privacy Filter detects and redacts PII in long-form text with a 96% F1 score, runs locally, and handles up to 128,000 tokens in one pass.
OpenAI has debuted Privacy Filter, a bidirectional token-classification model for detecting and redacting personally identifiable information (PII) that can scan long-form text in a single pass, run locally, and deliver greater context-awareness.
Scanning...
From vectors to tensors: Expanding the possibilities of AI search
In this live session, we will dive into how tensors extend what’s possible beyond traditional vector search. Join us on May 5 to uncover the sophisticated scoring needed for complex search problems involving multimodal data and nuanced relevance ranking.
Check out the updated author program at our sister site, Towards Data Science. Share your AI projects or tutorials with a massive audience and earn more with higher monthly caps and updated pay tiers.
As AI coding tools turn non-developers into builders, a new category of hyper-customized, economically viable software is emerging — and it's reshaping how teams work.
Developer-led observability: Debugging distributed and AI systems with runtime telemetry
Explore real-world examples and debugging demonstrations that show how to move faster, escalate less, and build more reliable AI-driven systems from the start.
