Most of us type things into AI apps we would never post publicly: a worry about a child, a draft of a hard email, a question about money. So it matters what happens next. We read each company's own help pages on 30 October 2026, and the news coverage of what has gone wrong, and put the answers side by side.
Who learns from you unless you say no
ChatGPT on Free, Plus and Pro, Gemini while "Keep Activity" is on, the free Microsoft Copilot app, Grok and Meta's new Muse agent all use your chats to improve their AI unless you switch it off. Claude asked every user to make the choice in 2025. Business and enterprise versions of ChatGPT and Copilot are off by default.
The switches: ChatGPT, Settings, Data controls, "Improve the model for everyone". Gemini, Settings and help, Activity, Keep Activity. Copilot, Privacy, "Training on conversation activity" and "Training on voice conversations". Claude, Settings, Privacy, "Help improve our AI models". Grok, Settings, Data Controls, "Improve the model". Muse, Settings, Data controls, "Help improve our AI models".
Who can read them
Google says a sample of Gemini chats is read by human reviewers, including outside contractors, and those chats are kept for up to three years even if you delete your activity. xAI says a limited number of its staff may read Grok conversations. Microsoft says it manually reviews some results. For ChatGPT and Meta AI, we could not confirm the answer from the companies' own pages.
How long they keep them
Deleted chats are removed within 30 days at OpenAI, Anthropic and xAI. Copilot keeps your history for 18 months, and Gemini auto-deletes after 18 months unless you change it. If you let Claude train on your chats, de-identified copies can be kept for up to five years.
The Share button is the biggest risk
In 2025 more than 370,000 shared Grok chats turned up in Google search, and people published private chats to the Meta AI app's public feed without realising. In July 2026 some shared Claude chats and Artifacts were found in Google search, some with health records and children's names. In October 2026 Meta paused sharing in Muse after about 6,800 shared conversations, including a loan application, were found exposed. Treat any share link as public.
Ads are arriving
Since 16 December 2025 Meta uses your chats with Meta AI to choose the posts and ads you see. ChatGPT's Free and Go plans now show ads chosen from what you are talking about. Google says Gemini chats are not used for ads, but from 16 October 2026 Gmail and Photos content you connect to its AI can help personalise ads in AI search.
What we suggest
Turn the training switch off on the apps you use. Use a temporary or private chat for anything sensitive. Keep health details, passwords and your children's details out of chats altogether. Do not press Share on anything you would not post. And look again in a few months: three of these companies changed their rules this autumn.