OpenAI offers zero retention for AI models
OpenAI is introducing zero retention for eligible API customers: it will not retain their instructions or responses after processing them, and it will not use their data to train models without explicit authorization. It is also testing Private Safety Processing, a system that detects risk patterns across multiple interactions without showing the content to OpenAI staff.
OpenAI has announced that it will offer zero data retention for eligible API customers: their instructions and model responses will not be stored after each request is processed. Business content will also not be used to train models unless the customer explicitly authorizes it.
The measure addresses a specific problem. Companies want to use advanced models with financial information, medical records, business plans, or private research, but many cannot allow that data to be stored or reviewed by the provider’s staff.
The challenge: detecting risks without storing content
Zero retention protects each interaction separately, but the most serious risks do not always appear in a single query. A series of seemingly normal requests could reveal an attempt to bypass safety measures, coordinate actions across multiple accounts, or keep an AI agent acting after the user has asked it to stop.
To address that problem, OpenAI is introducing Private Safety Processing in a testing phase. The system analyzes patterns across related interactions without allowing OpenAI staff to see the underlying instructions or responses.
The difference is that the automated system can study the context needed to identify potential abuse and send only a limited signal, such as the type of activity detected. That signal can help determine whether action is needed, but it does not include the original content.
OpenAI does not retain customer instructions or responses after processing the request.
Two ways to stay in control
With zero-retention deployments, content remains on infrastructure controlled by the customer. OpenAI is also developing an alternative that would store the data on its own infrastructure, but encrypt it with keys controlled by the customer.
Under that second model, OpenAI would not have a copy of the keys, so its staff could not access the underlying content. Automated systems would still be able to detect signs of potential misuse without exposing the conversations.
Companies could review alerts and enforcement decisions through their own systems. If they believe legitimate activity was flagged by mistake, want to file an appeal, or need to cooperate in an investigation into verified abuse, they could voluntarily share the relevant information with OpenAI.
It is not yet a general-purpose feature
Private Safety Processing is being tested with initial customers. OpenAI plans to begin rolling it out and publish a technical paper in September, although it has not yet detailed which customers will have access or the service’s full conditions.
The announcement matters especially for industries with strict privacy obligations. For example, a company could connect an assistant to internal documents or healthcare processes without accepting that every conversation would be available for human review by the provider.
For you, the change will depend on whether you use the OpenAI API through a company and on your account terms. It does not mean that every user automatically gets zero retention: it is an option for eligible customers and will still include automated safety systems.
The tension OpenAI is trying to resolve will remain central: models need more context to detect dangerous uses, but companies want that context to remain under their control. The next test will be to see, through the technical documentation and the September rollout, what signals are collected, how they are applied, and how far the promised privacy actually extends.