Conversations with ChatGPT, Gemini, Copilot and Claude are not always visible only to the user and the chatbot. The companies behind these services say certain interactions may be reviewed by people, including to improve their products, investigate potential policy violations or address safety concerns. This does not mean every chat is read by a human, but privacy settings and retention rules differ across platforms.
For ChatGPT, OpenAI lets users disable model training under Settings > Data Controls by turning off “Improve the model for everyone.” Temporary Chats are not used for training, but may be reviewed to monitor abuse. OpenAI says deleted conversations are generally scheduled for permanent removal within 30 days, subject to legal or security exceptions and information already de-identified.
Google says some Gemini conversations may be reviewed to improve services and protect users. Chats sent for review are disconnected from users’ accounts, but reviewed conversations may be retained for up to three years, even if the user deletes their activity. Turning off Keep Activity prevents future chats from being used to train Google’s models unless feedback is submitted, though conversations may still be kept for up to 72 hours for service and safety purposes, and some human review may occur.
Microsoft says some Copilot conversations undergo automated and human review for product improvement and digital safety. Chats suspected of breaching its Code of Conduct may also be examined, and users cannot fully opt out of review when suspected violations are investigated. Turning off the use of future eligible conversations for model training does not prevent other uses, such as security, compliance or certain product improvements.
Anthropic says conversations selected for model improvement are separated from account-identifying information before review, with access limited to a small group of personnel. Users can manage model-improvement settings in Claude’s privacy controls. Safety systems may still flag conversations for processing, while Incognito Chats are not used for general training and are generally retained for up to 30 days.
Disabling training and deleting a chat are not the same as preventing all access or erasing every stored copy. A paid subscription also does not automatically make a consumer chatbot confidential. Business-focused services may offer additional contractual protections and controls, but those differ by provider and product.
To reduce exposure, check each service’s privacy settings and use its temporary or incognito mode when appropriate. Avoid entering passwords, banking details, sensitive medical records or confidential work documents. When asking for help with sensitive material, replace names and other identifying details with placeholders, while remembering that context can still reveal someone’s identity.
For particularly sensitive tasks, locally run AI tools such as Ollama or LM Studio may reduce the need to upload information to a cloud service, provided they are configured to process it locally. Optional online features or network connections can still affect where data goes. These measures can limit exposure, but they do not guarantee that a conversation will never be reviewed for safety, security or legal reasons.
Sources: Tom's Guide.