What never to paste into an AI chatbot: an everyday privacy guide

In April 2023, Samsung engineers pasted internal source code into ChatGPT to get help with bugs. A few weeks later, the company banned the use of generative AI tools on its corporate devices. There was no attack and no technical failure: confidential information had simply left the company through copy and paste.

That is the most common risk with chatbots. It is not that AI "spies on you", but that what you type is no longer only on your device. Understanding where it goes lets you decide sensibly what to share.

What happens to what you type

When you send a message to ChatGPT, Claude or Gemini, the text travels to the provider's servers, where the model processes it. From there, three things come into play, depending on the service and your settings.

It is stored

Your conversations are saved so you can return to them. Even when you delete them or use temporary modes, providers usually keep them a few more days for safety reasons. OpenAI, for example, states in its help centre that temporary chats do not appear in your history, are not used for training and are deleted within 30 days.

It may be used to train models

On personal plans, the general rule is that your conversations may be used to improve models unless you switch it off. In August 2025, Anthropic changed its terms for Claude Free, Pro and Max users: since then, each person chooses whether their chats are used for training, and those who accept see the retention period for that data extended to five years. ChatGPT has an equivalent switch in its data controls.

Text being used for training does not mean the model will recite it to someone else, but it does mean it is no longer under your control.

A person may read it

To improve quality, some providers send a sample of conversations to human reviewers. The Gemini Apps privacy hub says so explicitly: it asks you not to enter confidential information you would not want a reviewer to see, and explains that reviewed conversations are kept for up to three years, disconnected from your account.

It may end up public

In mid-2025 it came to light that thousands of ChatGPT conversations shared by link were appearing in Google results. People had ticked a box allowing them to be indexed, often without understanding what that meant. OpenAI withdrew the feature, but the lesson still stands: a share link is a web page.

What you should never paste

These categories have one thing in common: if they leave your control, the damage is hard or impossible to reverse.

• Passwords, API keys and verification codes. No legitimate use of a chatbot needs them. If you are debugging code, replace them with text such as YOUR_KEY_HERE.
• Financial data. Card numbers, bank accounts, full statements, screenshots of your online banking.
• Identity documents. National ID, passport, social security number, driving licence.
• Health information. Medical records, diagnoses, lab results with your name on them, yours or someone else's.
• Other people's personal data. Full names with phone numbers, addresses or emails of customers, employees, patients or students. Here you are not only risking your own privacy: in many countries it may breach data protection law.
• Confidential company information. Source code, strategy, unsigned contracts, unpublished figures, anything covered by a non-disclosure agreement.
• Anything you would not say out loud in a coworking space. It is a good quick test for everything else.

How to use AI with sensitive information

The alternative is not to stop using AI, but to prepare what you give it.

1. Anonymise before you paste

Most tasks do not need the real data. To help you draft a reply to an upset customer, the model does not need their name or order number. Replace:

• Names with roles: "the customer", "the supplier", "Person A".
• Exact figures with rounded or proportional ones, if the result does not depend on the precise value.
• Companies and places with generic descriptions: "a mid-sized distributor in a Colombian city".
• Identifiers (contract, account or case numbers) with placeholders such as [CONTRACT].

Then, when you use the answer, you put the real data back yourself.

2. Share only the fragment you need

If you need a clause reviewed, paste the clause, not the whole contract. If you want an Excel formula, describe the columns and paste three made-up example rows. Less text also tends to produce more accurate answers (we explain why in Tokens and the context window).

3. Check files before uploading them

A document can carry more than you see: comments and tracked changes in Word, hidden tabs in a spreadsheet, metadata with the author's name, or layers in an image. Export a clean copy with only what you want to share.

4. Adjust your account settings

Spend five minutes reviewing the privacy options of the assistant you use:

• Training on your data. Find the option to stop your conversations being used to improve models and make a conscious decision.
• History. Consider whether you need to keep every conversation or whether temporary mode will do for the sensitive ones.
• Shared links. Review the list of conversations you have shared and delete the ones you no longer need.
• Connections to other apps. If you connected email, calendar or files, check what it can access and whether you still need it.

5. For professional work, use the right plan

If your work involves data about customers, patients or your company, personal plans are not the right place, even paid ones. Business plans and API access from the major providers include contractual commitments not to train on your data by default, and data processing agreements. If your organisation does not have one, ask before using an external tool with third-party information.

A different risk: what the AI reads for you

So far we have talked about what you give the chatbot. When you connect an assistant to your email or documents, or ask it to browse the web, another risk appears: malicious text inside that content giving the model instructions. It is called prompt injection and we explain it in detail in Prompt injection: the risk of AI obeying someone else.

Quick checklist before you paste

Before you send, ask yourself:
- Are there passwords, keys or codes? Remove them.
- Are there other people's names, phone numbers, emails or documents? Anonymise.
- Is there information my company or a client would consider confidential?
  Use a business plan or do not share it.
- Does the model need all this text, or is a fragment enough?
- If this chat were published tomorrow, would it cause me a problem?

Keep reading

• Prompt injection: the risk of AI obeying someone else: what happens when you connect AI to your email or the web.
• How to write a prompt that works the first time: how to give useful context without real data.
• How to summarise long documents with AI: techniques for working with reports and contracts.
• Claude skills, connectors and plugins: what connecting your apps to an assistant involves.
• In the glossary: data leakage, prompt injection and guardrails.

Sources

• Anthropic. Updates to Consumer Terms and Privacy Policy. August 2025.
• OpenAI. Data Controls FAQ. OpenAI Help Center.
• Google. Gemini Apps Privacy Hub. Gemini Help.
• The Japan Times (with Bloomberg reporting). Coverage of Samsung's generative AI ban. 2 May 2023.
• TechRadar. OpenAI pulls chat sharing tool after Google search privacy scare. August 2025.