OpenAI has announced a new system, Private Safety Processing, designed to enhance the detection of AI misuse while maintaining user privacy. The company is currently previewing this system with select early customers and plans a broader rollout, along with the publication of a technical white paper, in September.
Private Safety Processing aims to identify patterns of potential misuse across related user interactions without allowing OpenAI personnel to access the underlying content of those interactions. This builds upon existing automated safety measures that typically evaluate requests individually.
For eligible API customers utilizing Zero Data Retention (ZDR), prompts and model responses are not retained after processing. An exception is made for images flagged as potential Child Sexual Abuse Material (CSAM), which may be retained for manual review and reporting to authorities. OpenAI has also stated that enterprise customer data is not used for model training unless customers explicitly opt in.
The ZDR deployment model involves customers retaining content on their own infrastructure. OpenAI is also developing an alternative where content would be stored on the company's infrastructure, secured with customer-controlled encryption keys. In both configurations, automated systems are designed to identify potential misuse and transmit limited safety signals without revealing the actual prompts or responses.
When the system detects a potential risk, OpenAI receives a defined signal indicating the nature of the activity. This information can then inform enforcement decisions. Customers have the ability to investigate alerts using data within their own systems and can share relevant details with OpenAI to appeal a decision, clarify legitimate activities, or support investigations into verified abuse.
OpenAI emphasized that addressing emerging risks in AI requires collaborative efforts, and Private Safety Processing has been shaped by input from customers across various industries, regions, and company sizes. The company highlighted the importance of protecting sensitive data and ensuring accuracy and integrity, particularly in sectors like healthcare, to build trust with users.
