OpenAI has announced a new "Privacy Filter" feature that will automatically redact or block prompts containing what the company identifies as personally identifiable information (PII) before they reach its models [source].

The filter, rolling out across ChatGPT and API endpoints, scans user inputs for patterns matching email addresses, phone numbers, credit card numbers, and other sensitive data. When detected, the system either replaces the information with placeholder tokens or rejects the request entirely, depending on configuration settings.

OpenAI states the feature aims to reduce accidental data exposure and align with enterprise compliance requirements. API customers can toggle the filter on or off via dashboard settings, while ChatGPT users will see it enabled by default with no opt-out for free-tier accounts.

The announcement does not specify which detection methods are used, whether the scanning occurs client-side or server-side, or how the system handles edge cases such as obfuscated data or non-English PII formats. OpenAI notes that the filter may produce false positives and advises users to review outputs for unintended redactions.

The feature arrives amid ongoing scrutiny of how AI providers handle user data. OpenAI has previously faced questions about data retention policies and whether prompts are used for model training. The company states that filtered content is not stored or logged, but the announcement does not detail how the filter itself is audited or whether metadata about blocked requests is retained.

Privacy Filter is available immediately for ChatGPT Plus, Team, and Enterprise users, with API access in public beta. OpenAI has not disclosed whether the feature will be extended to free-tier API users or older model versions.