Are AI chatbots safe? Privacy of ChatGPT, Claude, Gemini and others
Are AI chatbots safe to use? It depends less on the chatbot than on what happens to what you type.
| Name | Made by | Company based in | Trains on your chats | People can read chats | Temporary chat mode | Chats used for ads | Minimum age | Regulator actions listed |
|---|---|---|---|---|---|---|---|---|
| ChatGPT | OpenAI | United States | Yes, unless you opt out | Yes | Yes | Yes | 13 | 7 |
| Claude | Anthropic | United States | You choose when prompted | Limited | Yes | No | 18 | 0 |
| DeepSeek | DeepSeek | China | Yes, unless you opt out | Not stated | No | No | 18 | 15 |
| Gemini | United States | Yes, unless you opt out | Yes | Yes | No | 13 | 2 | |
| Grok | xAI | United States | Yes, unless you opt out | Yes | Yes | Not stated | 13 | 4 |
| Meta AI | Meta | United States | Yes, no opt-out for chats | Yes | No | Yes | 13 | 2 |
| Microsoft Copilot | Microsoft | United States | No | Yes | No | Yes | 13 | 0 |
| Mistral Vibe (formerly Le Chat) | Mistral AI | France | Yes, unless you opt out | Limited | No | Not stated | 13 | 1 |
| Perplexity | Perplexity | United States | Yes, unless you opt out | Not stated | Yes | No | 13 | 1 |
This table compares nine AI assistants on the questions that decide that: whether your chats train the company’s models by default, whether people can read them, how long they are kept, whether they are used for ads, where the data is stored, and what regulators have done. Every fact comes from the company’s own privacy policy, terms and help pages, or from a regulator’s own website, and each assistant’s page links to them.
Checked on September 30, 2026. Policies change often, so each page shows the date we last checked it.
What the table shows
- Trains on your chats. “Yes, unless you opt out” means personal-plan chats are used to train or improve models until you change a setting. Business and API plans usually have different terms; each page describes them.
- Human review. Whether employees or contractors can read conversations, as the company describes it.
- Temporary chat. A mode that keeps chats out of your history and out of training, usually deleted within a set time.
- Ads. Whether chat content is used to choose or personalise ads.
- Regulator actions. Orders, fines, bans and formal inquiries announced on regulators’ own sites. An inquiry is not a finding of wrongdoing.
Three things worth doing on any AI assistant
- Check the training setting. On ChatGPT, Gemini, Grok, DeepSeek, Perplexity and Mistral’s Vibe, personal chats are used for training until you turn it off. Feedback you give with thumbs up or down can be used even after you opt out.
- Use temporary or incognito chats for sensitive questions. ChatGPT, Claude, Gemini, Grok and Perplexity offer one.
- Don’t paste what you wouldn’t want a reviewer to see. Most companies reserve the right to have people read some conversations, for safety or to improve the product.
For what happens to documents you upload, read what happens to your files when you upload them to an AI tool. For the difference between running a model on your own machine and in the cloud, see on-device vs. cloud processing.
Questions
Which AI chatbot is the most private? None of them is private by default in every respect. On the facts in this table, Claude asks you to choose whether chats train its models and has no ads, Copilot’s updated app says chats are not used to train its foundation models, and Mistral hosts data in the EU by default. Each still keeps chats until you delete them and can review some of them.
Are AI chats used for advertising? Meta and Microsoft use AI chats to personalise ads, and ChatGPT’s Free and Go plans show ads chosen from the conversation. Claude has no ads, Google says Gemini chats are not used to show ads, and DeepSeek and Perplexity say they do not use conversations for targeted advertising.
Does deleting a chat delete it everywhere? Usually within 30 days, but not always: Gemini keeps chats a human reviewer has seen for up to three years, and several companies keep data longer for legal or safety reasons.
Are business plans different? Yes. ChatGPT Business and Enterprise, Claude for Work, Google Workspace, Copilot for work accounts, Grok business plans and Perplexity Enterprise are not used for training by default.
Sources
Each assistant’s page lists the company pages and regulator decisions it is based on, with the date checked.