Can people at OpenAI, Google and Microsoft read your AI chats? What you should know about privacy

AI chatbots are used for personal advice and work documents. But conversations with ChatGPT, Gemini, Microsoft Copilot and Claude may not be visible only to the person who typed them. Their providers allow selected chats to be reviewed by people for safety, troubleshooting and AI improvement.

That does not mean every conversation is routinely read. Automated systems can flag harmful requests, suspicious activity or policy violations for further checks. Some chats may also be assessed to improve AI responses, depending on the service and user settings. A Tom’s Guide report examined these practices across AI platforms.

MUST READ | IIT students study Sanskrit and shastras to strengthen reasoning skills, says IISc professor

Turning off AI training does not stop every review

In ChatGPT, you can disable “Improve the model for everyone” under Settings > Data Controls. Temporary Chats are excluded from model training, but may be reviewed for abuse and are generally deleted after 30 days.

On Gemini, switching off Keep Activity prevents future chats from being used to train Google’s models, unless you submit feedback. Conversations may still be retained for up to 72 hours for service and safety purposes. Chats reviewed by people can remain stored for up to three years, even after activity is deleted.

DON'T MISS | AI is changing how you work, but are you learning to use it effectively?

In Copilot, you can switch off “Training on conversation activity” under Profile settings > Privacy. Microsoft says some human review may still take place, including investigations into suspected Code of Conduct violations. In Claude, you can manage model-improvement permissions under Settings > Privacy. Incognito Chats are excluded from general training but may be retained for around 30 days.

Deleting chats or paying for AI may not be enough

Deleting a ChatGPT conversation schedules its removal within 30 days, subject to legal, security and de-identification exceptions. A paid consumer subscription does not automatically make chats confidential. Business and enterprise services may offer contractual protections, but these vary.

Avoid sharing passwords, banking details, sensitive medical records or confidential work files with consumer chatbots. Removing names and account numbers can help, but context may still identify someone. For sensitive tasks, locally run AI tools can reduce the need to send information to cloud services, provided online features are disabled.