AI and privacy: what not to type and which settings to check

Where does your text go?

When you type a question into an AI service such as /en/tools/chatgpt/, /en/tools/gemini/ or /en/tools/claude/, the text doesn't stay on your own device. It travels over the internet to the company's servers, where the AI works out its answer. The conversation is usually saved to your account too, so you can come back to it later.

That isn't dangerous in itself – email and online banking work the same way. The difference is that people tend to tell an AI things they would never write anywhere else. The chat feels private, even though there is a company's server on the other end.

On top of that, many free and consumer services may use your conversations to improve – or "train" – the AI unless you opt out in the settings. Training means the texts are used as learning material so the AI gets better at answering. In some services, people may also review snippets of conversations for quality checks.

Don't type these into an AI

A good rule of thumb: don't tell an AI anything you wouldn't want to see in a stranger's inbox. In particular, leave out:

• Passwords, PINs and banking codes. No genuine AI service needs them.
• Your national ID number, passport or driving licence number, and bank account numbers.
• Other people's health details and sensitive matters, like a neighbour's illness or a grandchild's school support plan, with names.
• Personal data about customers and staff: names, phone numbers, addresses, salaries.
• Business secrets: quote prices, contracts, inventions, future plans.
• Photos of documents showing an ID number, address or account number.

You can of course ask about your own matters. If you'd like help understanding a lab result, type just the values and your question – not your name, ID number or patient number. There's more on this in /en/guides/health-information-wisely/.

Settings worth checking right away

Most services let you decide whether your chats can be used for training and whether your history is kept. Menu names change from time to time, but you'll usually find these under Settings, Data controls or Privacy.

• ChatGPT: Settings → Data controls → turn off the option that lets your chats improve the model for everyone. You can also use temporary chats, which aren't saved to your history, and check the memory setting, which lets the service remember things about you.
• Gemini: In your Google account's activity settings you can switch off saving for Gemini apps and delete old conversations. Temporary chats are available too.
• Claude: Settings → Privacy → choose whether your chats can be used to improve Claude.
• Copilot: The settings include a privacy section where you can opt out of model training on text and voice.

Check these once on every device you use. When you're signed in with the same account the choice usually carries over, but it's worth double-checking.

Small businesses: customer data and trade secrets

For small business owners, AI is a handy helper for quotes, social media posts and emails. But more is at stake with business data: your customers' personal data is protected by the EU's General Data Protection Regulation (GDPR), and you are responsible for how it's handled.

1. Use a business version for work. For example, /en/tools/m365-copilot/ with a work account and the business or team plans of other services generally don't use customer data for training, and they come with a data processing agreement. Always confirm this in the terms.
2. Agree on house rules. Write down briefly what staff may use AI for and what must never go in. One page is enough.
3. Strip identifiers before pasting. If you want to summarise customer feedback, remove names and contact details first.
4. Don't let AI make decisions about people. In hiring or credit decisions, a human has the final say. More in /en/guides/ai-and-the-law/.

Ideally use this prompt in your company's own secure version – or at least after you've removed the most obvious details yourself.

Attachments, images and voice

Many services accept PDFs, photos and speech. The same rule applies: everything you attach goes to the server. A screenshot may accidentally show email addresses, browser tabs or customer details. Crop the image before you send it.

When you talk to an AI by voice, your speech is turned into text and may be stored just like a typed chat. Never read out a card number or codes.

How to spot a trustworthy service

There are hundreds of AI services out there, and not all of them have a clear background. Signs of a trustworthy one:

• a well-known company stands behind it, with easy-to-find contact details and a privacy policy
• you download the app from an official app store and the developer name matches (see /en/guides/ai-on-your-phone/)
• the settings let you manage your data and delete conversations
• it doesn't ask for unnecessary permissions, such as your contacts or text messages.

Be suspicious if a site promises "free premium AI" in exchange for logging in with Facebook, or asks for card details before you've seen anything. Read more about scams like these in /en/guides/ai-scams-and-deepfakes/.

Safe-use checklist

1. Use a strong password you don't use anywhere else, and turn on two-step verification.
2. Check the training and history settings in every service you use.
3. Don't type passwords, ID numbers, account numbers or other people's health details.
4. Remove names and contact details before pasting text or images.
5. For work tasks, use a version your company has approved.
6. Double-check important answers – AI can be wrong while sounding certain.
7. Delete old conversations regularly if you don't need them.
8. Sign out of shared computers, such as those in a library.

Once these are in place, you can use AI with peace of mind. If you're just getting started, carry on to /en/guides/get-started-with-ai/.