What happened
Zhipu AI has announced a new privacy-focused for its Model-as-a-Service (MaaS) platform that will prevent the static storage of user input and output data following model . According to the report from AIBase, this mechanism ensures that data is discarded immediately after a model call is completed, rather than being retained on the platform's servers.
Zhipu AI's new mechanism is designed to ensure that data generated during model calls is not stored statically on the platform. The company describes this as a 'data content not stored' , where information is used solely to fulfill the immediate request and is subsequently discarded.
The is intended to serve as an optional layer of 'data insurance' for enterprise customers who have strict data compliance requirements. Interested users can apply for access through the Zhipu MaaS console, though the platform has not yet finalized the specific deployment schedule or the full scope of the feature's availability.
The policy includes significant caveats. Zhipu clarified that the mechanism does not apply to all services; specifically, the Batch API and File API are excluded because they require persistent storage to function. Furthermore, the platform reserves the right to retain data for 30 days or longer to satisfy legal and regulatory obligations or to investigate potential platform abuse and security violations.
Source details: news.aibase.com ↗
Why it matters
This development addresses critical enterprise concerns regarding data sovereignty and compliance when deploying large language models. By offering a 'data non-preservation' option, Zhipu provides a mechanism for organizations to mitigate risks associated with long-term data retention. However, the utility of this is limited by necessary exceptions for regulatory compliance and specific API functionalities, highlighting the ongoing tension between privacy-preserving AI and the operational requirements of model platforms.
Data privacy is a primary barrier to the adoption of large AI models in enterprise environments. By introducing a mechanism that effectively 'burns' data after use, Zhipu is attempting to provide a compromise that respects user data sovereignty while maintaining the operational efficiency of its MaaS platform.
The announcement underscores the complexity of implementing 'zero-retention' policies in AI. Because platforms must remain accountable for security and legal compliance, a truly absolute 'no storage' policy is rarely feasible. Zhipu's approach acknowledges this reality by explicitly carving out exceptions for abuse monitoring and regulatory requirements, which are essential for maintaining a secure and compliant AI ecosystem.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which of these is a common misconception about AI Ethics?
What to watch next
The platform has not yet provided a specific rollout date, noting that implementation timelines and the exact scope of the will be determined by the platform. Users are currently invited to submit applications for the feature via the MaaS console. It remains to be seen how Zhipu will balance this 'no storage' policy with its stated requirement to retain data for at least 30 days to monitor for abuse or to comply with legal mandates.
The primary uncertainty lies in the actual implementation timeline, as Zhipu has not provided a concrete date for when this will be active for applicants.
Observers should monitor how the platform defines the 'applicable scope' of this , as the current announcement leaves significant ambiguity regarding which specific models or API endpoints will be eligible for the non-preservation status.
The interaction between this and the 30-day retention period for abuse monitoring will be a key point of interest for privacy advocates and enterprise security teams, as it defines the actual boundary of the 'non-preservation' promise.