AI assistants collect user data for model training and safety reviews
OpenAI and Google retain temporary chats for specific periods, and users can adjust settings to limit how their personal information is processed and stored.
By Marco Kapoor ·
Consumers increasingly rely on artificial intelligence tools for personal and professional tasks, but the data exchanged in these interactions is often stored and analyzed by providers. As usage expands, the volume of personal information passed to these systems grows, creating a persistent digital footprint that users may not fully understand.
The primary risk involves the retention and utilization of user data. Providers store conversations, uploaded documents, and audio recordings. This data may be reviewed by human employees or used to improve the underlying AI models. In some cases, chatbot data can be used to build user profiles or disclosed to other users, depending on the specific service and its configuration.
OpenAI and Google, the developers of ChatGPT and Gemini respectively, outline distinct policies for handling this information. OpenAI states that temporary chats are not used to improve its models but are retained for up to 30 days for safety purposes. Google indicates that Gemini temporary chats are not used for model training but are retained with the user's account for 72 hours.
Data collection extends beyond typed text. Google's Gemini privacy notice notes the collection of device and location data, as well as information from connected services. Linking an AI assistant to an email account or sharing a screen expands the range of data the system can process. The Federal Trade Commission advises users to review permissions for microphones, cameras, photos, contacts, and location to limit this access.
Users can reduce data retention by adjusting specific settings within their accounts. For standard personal use of ChatGPT, individuals can disable the option to improve the model for everyone under the data controls menu. Saved chats remain accessible, but voice sharing and feedback submissions may still allow associated conversations to be used for training if those specific controls are active.
Deleting a chat does not always remove all associated data. OpenAI explains that a separate saved memory may persist even after a conversation is deleted. To fully remove remembered information, users may need to delete the memory, the relevant chats, and other sources where the data appears. Turning off the memory feature does not automatically delete existing conversations.
Deletion policies have further limitations. Google states that some previously reviewed chats remain for up to three years, disconnected from the user's account, even after activity is deleted. This retention period applies to data that has already been processed by the system, ensuring it is no longer linked to the active account but is not immediately erased from provider servers.
Karen Clay, who operates Clay Technology and Multimedia, recommends that users provide AI tools with the problem they need solved without providing their full identity. She advises replacing names with labels like Person A and removing identifying details from documents before uploading them. A distinctive combination of employer, location, and personal circumstances can identify an individual even without a name.
Before sharing sensitive information, users should check the current privacy policy of the specific tool they are using. Settings and exceptions vary by service, account type, and feature. The FTC also suggests using physical microphone switches on smart speakers during sensitive conversations and protecting accounts with unique passwords and multifactor authentication.
