OpenAI Introduces Privacy-Focused Monitoring Service Amid Competition with Anthropic

As the capabilities of AI models expand, so too does their potential for misuse, prompting calls for enhanced safety measures. In response, OpenAI has launched a new service called Private Safety Processing, aimed at monitoring misuse while ensuring customer data remains private.

This new system, which is currently being previewed to select clients, differs significantly from a recent data retention policy implemented by OpenAI’s competitor, Anthropic. Anthropic’s policy allows them to retain user data, including all sessions and conversations, for 30 days relating to specific “covered models,” including all Mythos-class models. This policy, introduced in July, aims to enhance safety by enabling the company to analyze potential abuse, but it has raised concerns among clients handling sensitive data.

OpenAI, adhering to a Zero Data Retention (ZDR) policy, allows clients to monitor abuse without retaining any user data. This pre-existing framework enables detection of misuse at a session level without human oversight. While Anthropic also follows a ZDR policy generally, exceptions exist for “covered models” like Fable.

The Private Safety Processing technology expands the ZDR concept by providing long-term monitoring capabilities across multiple user interactions. OpenAI states that this system can track activities over time, identifying patterns indicative of misuse without retaining the content of conversations. For instance, if a malicious individual were to spread out requests to evade detection, Private Safety Processing could flag such behavior based on cumulative interactions.

When suspicions are raised, the system may generate a specific alert signal for OpenAI, allowing the company to assess whether further enforcement actions are necessary. If a situation arises, OpenAI may contact the client for additional context, although sharing any user data remains at the client’s discretion.

In contrast, Anthropic permits human reviews of customer data under strict controlled access, which involves a limited number of approved reviewers. Such reviews are documented in a secure log that cannot be altered or erased.

The competitive landscape between OpenAI and Anthropic remains intense, with both firms vying for market advantage. Recent reports indicate that OpenAI’s growth in the second quarter has lagged behind Anthropic’s, as the latter’s annual revenue run rate reportedly reaches $65 billion, with speculation of a potential IPO valuation as high as $2 trillion.