What to tell AI and what to keep private
AI needs context, not your entire digital history. A short guide to what to share, what to remove, and what to check before connecting your accounts.
AI needs context, not your entire digital history. To get a useful answer, explain the task. Add personal or confidential information only if it is genuinely needed and you are allowed to share it.
Illustration: Synesis, AI-generated.
Describe the problem, not the person
Four things help most: what you want to achieve, who the answer is for, the relevant circumstances, and the constraints.
Instead of pasting an entire email with a signature, phone number, and order details, write:
A customer received a damaged product. We can offer a replacement or a refund. Draft a short, friendly reply without making excuses. Do not promise a delivery date because we do not know it yet.
For this draft, AI does not need the customer's name or address. You can add those details yourself once the text is ready.
Remove unnecessary details before pasting
Replace names with labels such as “customer A” or “employee B”. Leave out contact details, identification numbers, and anything else that does not affect the task. If you need help with one paragraph of a contract, do not automatically send the whole contract.
Removing a name is not always enough. A combination of someone's job, location, and the event described can also identify them.
With screenshots, check the background too: open tabs, notifications, addresses, and other information you did not intend to share.
Keep some things out of the chat
Do not paste passwords, login codes, recovery codes, or access keys. If you need technical help, replace them with [REMOVED].
Pause before sharing medical reports, bank statements, personnel documents, or trade secrets. Check whether the tool is suitable for that information and whether you have permission to share it. For work documents, use a tool and account your organisation has approved for that purpose.
Being able to open a document does not mean you are allowed to upload it to an external service.
Check the settings, not just the subscription name
Chat history, assistant memory, and using conversations to train models are different things. ChatGPT, for example, lets you turn off the use of conversations for model improvement while keeping those conversations in your history.[1]
Check:
- Does the provider use your content to train models?
- How long is it stored, and how can you delete it?
- What does the assistant remember for future conversations?
Do not assume that “paid” means suitable for all confidential information, or that “temporary chat” means no retention at all. OpenAI, for example, states a 30-day retention period for temporary chats.[1]
Be more selective when connecting accounts
Pasting one paragraph is not the same as connecting your inbox or a folder of documents. Before connecting anything, check what the assistant will be able to access and how to revoke that access.
Separate permission to read from permission to act. If you want an email summary, you do not need automatic sending of replies. Grant the smallest scope of access that still supports the task.
Before clicking “Send”, ask three questions: Does AI need this information? Am I allowed to share it? Can I complete the task with less data?
A good description of the task is usually more useful than another attachment.