OpenAI has published a new explanation of how ChatGPT develops broad knowledge while attempting to limit the use and exposure of personal information. The post is framed as a plain-language guide to model training, privacy safeguards, and the controls available to people using ChatGPT.
The company says ChatGPT's growing ability to help with coding, research, analysis, and multi-step real-world tasks depends on training over a wide range of information sources. But it also acknowledges that building more capable models creates a responsibility to reduce personal information in training, give users clearer controls, and keep privacy protections aligned with safety requirements.
What information may be used in trainingAccording to OpenAI, the models that power ChatGPT are developed using a mix of publicly available information, information accessed through partnerships, and information provided or generated by users, contractors, and researchers. The company says that mixture helps models build general knowledge and respond more reliably and safely.
For publicly accessible internet content, OpenAI says it uses information that is freely and openly available. It gives examples such as public forum posts, blogs, and other publicly published material. The company also says it may use third-party and partnership data where permitted, while maintaining that these sources are meant to support model quality rather than collect personal information for its own sake.
How OpenAI says it reduces personal informationOpenAI says it uses an internal version of Privacy Filter at multiple stages of the training process. That includes public datasets used for training and user conversations when the account setting Improve the model for everyone is enabled. The stated purpose is to reduce personally identifiable information before it is used in training workflows.
The company also notes that it has made Privacy Filter available to other developers at no cost, positioning the tool as part of a broader effort to reduce privacy risk across the industry rather than only within OpenAI's own systems.
Privacy controls inside ChatGPTOne of the core messages in the post is that users can decide whether their conversations contribute to future model training. OpenAI says people can turn off Improve the model for everyone in settings under Data Controls. When that setting is disabled, new conversations remain visible in chat history but are not used to improve ChatGPT.
The company also points to Temporary Chat as a separate option. Temporary chats do not appear in chat history, do not create memories, and are not used to improve the models. OpenAI says those conversations are retained for 30 days for safety purposes and then deleted.
Privacy in model responsesOpenAI says ChatGPT is designed to reject requests for private or sensitive information about individuals. At the same time, the company acknowledges that the system can make mistakes. If someone believes a ChatGPT response includes personal information that is inaccurate or inappropriate, OpenAI says they can submit a request through the privacy request portal.
That part of the policy matters because privacy protections are not only about what goes into training. They are also about how the model behaves at inference time, how it handles requests about individuals, and what recourse people have when something goes wrong.
A policy signal as much as a product explanationThe post is also a policy document in practice. OpenAI is trying to show that model capability, privacy safeguards, and user controls can be explained together in a way that is easier for non-specialists to understand. That is increasingly important as people use ChatGPT for more personal and professionally sensitive tasks.
The company says protecting privacy and addressing serious risks of harm have to work together, not compete with one another. As models become more capable, OpenAI says it plans to keep improving safeguards, make privacy controls clearer, and give users more concrete ways to decide how their information is used.
Official source: OpenAI.