OpenAI introduces new privacy measure to prevent AI misuse
What's the story
OpenAI has unveiled a new privacy-focused security measure, called Private Safety Processing, to prevent the misuse of its artificial intelligence (AI) models. The automated system detects potential abuse without storing any customer data. The move comes as a direct response to rival company Anthropic's recently announced data retention policy for covered models that allows user data storage for up to 30 days.
Policy backlash
Anthropic's controversial data retention policy
Anthropic's data retention policy, announced in July, allows the AI lab to store user data for 30 days for "covered models." These include all Mythos-class models and future ones with similar capabilities.
The move was intended to enhance safety by enabling the lab to detect potential impropriety.
However, it has raised privacy concerns among enterprises that handle sensitive data and don't want their information stored or inspected by an AI lab.
Policy extension
OpenAI already had a 0 data retention policy
OpenAI, like many other AI companies, already offers a certain level of privacy with its Zero Data Retention (ZDR) policy.
ZDR uses agents within the OpenAI API to monitor abuse on a per-session basis without retaining customer data.
Anthropic also follows ZDR but not for "covered models," such as Fable.
OpenAI says Private Safety Processing is an advanced version of ZDR that monitors inputs and outputs across multiple conversations for signs of potential misuse.
Detection mechanism
What is private safety processing?
The new technology, Private Safety Processing, helps OpenAI detect malicious AI use across multiple sessions.
This could be a case where a bad actor spreads out their requests to avoid detection.
The system can analyze these conversations for signs of abuse without human review of user conversations. If triggered, it sends a "narrowly defined signal" to OpenAI warning about specific activity types.