What you should never tell ChatGPT (or Gemini, or Claude): an AI privacy guide
Passwords, ID numbers, medical records, photos of other people… Here's what to keep out of AI chatbots, how to stop your chats from training AI in ChatGPT, Gemini and Claude, and what privacy regulators recommend.

We ask AI about back pain, paste in a lease to get a summary, or upload a photo of our blood test results. It’s convenient. But everything you type into a chatbot leaves your phone and is stored on a company’s servers.
There’s no need to be scared of AI. You just need to know what to keep to yourself and which settings to change. This guide draws on advice from Spain’s data protection agency (AEPD) and its national cybersecurity institute (INCIBE), which apply just as well wherever you live.
At a glance
| Golden rule | Don’t type anything you wouldn’t put in an email to a stranger |
| Never share | Passwords, codes, bank details, ID numbers, medical data with your name |
| Handy trick | Swap real names and details for made-up ones: the AI helps just as well |
| Key setting | Turn off using your chats to train the AI |
| For sensitive topics | Use a temporary or incognito chat |
Why it matters what you share
- Your chats are stored. Conversations stay in your account and on the company’s servers for some time.
- They may train the AI. On many services this is switched on by default, as INCIBE points out.
- Deleting isn’t instant. Removing a chat from your history doesn’t guarantee it disappears right away from the provider’s internal systems.
- People may read them. Some conversations are reviewed for safety or to improve the product.
- Accounts get hacked. If someone gets into your account, they can read your whole history.
What you should never type into a chatbot
| Type of data | Examples | What to do instead |
|---|---|---|
| Logins | Passwords, text-message codes, PINs, recovery phrases | Never. No chatbot needs them |
| Bank details | Card numbers, account numbers, screenshots of your banking app | Ask with round or made-up numbers |
| ID documents | Passport, driver’s license, Social Security or national ID number | Cover them before uploading a photo or PDF |
| Health data with your name | Medical reports, test results, diagnoses | Remove your name and record numbers; always check with your doctor |
| Other people’s data | Photos of friends or kids, client or coworker details | Don’t upload them without permission, especially minors |
| Work secrets | Contracts, internal figures, company code | Only use the AI your employer has approved |
| Where you are | Your exact address, routines, live location | Say “a mid-sized city” or “a neighborhood” |
The AEPD’s advice boils down to this: don’t share personal data (full name, address, phone number, ID numbers) or sensitive information (medical, financial or contractual). If you need help with a specific situation, describe a made-up case that looks like yours.
A warning about photos: the AEPD notes that using images of other people to create new content, especially minors, can break data protection law and may even be a crime.
The trick: anonymize before you ask
The AI doesn’t need to know who you are to help you. Compare:
- ❌ “I’m Laura Smith, I live at 12 Main Street and my bank charges me a $15 fee on account 0012…”
- ✅ “My bank charges me a $15 monthly fee. How can I avoid paying it?”
The answer will be just as useful. Before pasting a document, delete or black out names, account numbers, addresses and signatures.
How to stop your chats from training AI
Menus change often, but this is roughly where things are today:
| Service | Where to turn off training | “No trace” mode |
|---|---|---|
| ChatGPT | Settings → Data controls → “Improve the model for everyone” | Temporary chat: not saved to your history or used for training |
| Gemini | Gemini Apps Activity (Keep Activity). Google is splitting history and training into two separate settings | Turn off activity (you lose your history) |
| Claude | Settings → Privacy → “Model Improvement” | Incognito chats: never used for training, even if the setting is on |
A few important details:
- ChatGPT: according to 2026 privacy guides, even with temporary chats or history turned off, OpenAI keeps conversations for about 30 days for safety monitoring before deleting them.
- Claude: Anthropic says that if you give a response a thumbs up or down, the whole conversation may be kept for up to 5 years (de-linked from your user ID) and used for training. If privacy matters, don’t rate chats that contain sensitive details.
- Gemini: until now, one setting controlled both whether your history was saved and whether it trained the AI. We explain the upcoming change in our article on Gemini Tasks and the new privacy setting.
- Copilot, Meta AI and others: look in their settings for words like “privacy”, “activity” or “training”. If you can’t find anything, be even more careful about what you type.
Watch out for memory and AI agents
Assistants now remember things about you across chats, and some can act for you: read your email, shop, or call a business. In October 2026 the AEPD published a checklist for using them wisely. The main ideas:
- Ask yourself five questions: what can it see, what can it remember, what can it do in my name, where does it send my data, and does it affect other people?
- Grant only the permissions each task needs. Avoid “always allow” unless you understand what it means.
- Require confirmation before it sends, shares, pays or signs up for anything.
- Check its memory in the service’s settings and delete what you don’t want it to keep. Telling it “forget this” in the chat isn’t enough.
- Beware of hidden instructions in emails or websites: someone can hide commands meant to trick the agent.
- At work, only use approved tools. Don’t connect your work email to a personal AI account.
Protect your account too
- Use a long, unique password for your AI account.
- Turn on two-step verification.
- Give the app only the permissions it needs (microphone, camera, files).
- Review your privacy settings every few months: companies change them.
Health, money and feelings: AI is not a professional
AI can help you understand a medical term or organize your spending. But as the AEPD reminds us, even when it seems to understand you, it doesn’t actually understand or empathize. For a diagnosis, a big financial decision or a rough emotional patch, talk to a qualified professional. And always double-check what it says: AI sometimes makes things up.
If you have kids
The AEPD asks parents and guardians to explain to children what they should never tell an AI, and to stay involved. If you share a device, set up separate profiles.
What if I’ve already shared something I shouldn’t have?
- Delete that conversation from your history.
- If it was a password, change it now.
- If it was bank details, contact your bank.
- Turn off training for everything you type from now on.
- If you think someone has misused your data, report it to your local consumer protection or cybercrime authority (in the US, for example, the FTC at reportfraud.ftc.gov).
Bottom line
AI is safe to use if you treat it like a very smart stranger: ask it anything, but don’t hand over your keys. Strip out personal details, switch off training, and use temporary mode for anything sensitive.
Sources
- Spanish Data Protection Agency (AEPD): «Careful what you confide» checklist (January 27, 2026)
- AEPD: checklist for delegating tasks to AI assistants and agents (October 27, 2026)
- INCIBE (Spain's cybersecurity institute): privacy settings for AI tools (September 16, 2026)
- Anthropic: Is my data used for model training? (March 16, 2026)
- Analytics Vidhya: AI chatbot privacy guide (June 2026)

