This is when not to use AI because the input itself is too sensitive

This is when not to use AI because the input itself is too sensitive.

Most safety advice focuses on the answer: verify the price, check the warranty, ask a professional before acting. Sometimes the bigger risk happens before the answer. The task seems useful only if you hand an AI service a medical record, identity document, private conversation, financial statement, confidential work file, or a detailed view into your home.

That is not one generic warning attached to every upload. It is a boundary you can decide in a few seconds by asking whether the sensitive information is incidental or whether it is the thing the AI would need to understand.

Three kinds of input

1. Ordinary. Public product pages, a menu, a manufacturer manual, a harmless preference, or a tight photo of an appliance display. Use the minimum the task needs and check the result.
2. Selective. The useful detail is ordinary, but private material may appear around it. Move personal objects before taking the photo, physically cover identifiers, frame only the needed rows or object, and inspect the newly created capture before uploading. If that takes more than a quick, obvious step, stop.
3. Sensitive at the center. The diagnosis, account history, identity, private message, confidential terms, child’s information, or home layout is the useful input. A smaller capture would remove what makes the task work. Do not upload it.

A photo is not automatically safer

Photographing one part of a page can be a good selective-capture technique only when the sensitive material is incidental. Cover the name and account details first, take a new tight image of the harmless rows, and look at the image before sending it. Photographing an entire sensitive document and hoping the service ignores the rest is still sharing the entire document.

The same rule applies beyond paper. A screenshot can reveal notifications and other tabs. A room photo can reveal faces, mail, medication, security equipment, possessions, and layout. A spreadsheet can remain confidential even after names are removed because the numbers and combinations identify the business or person.

The hard-stop list

• Medical records, lab results, portal screens, and detailed personal health histories
• Government IDs, tax documents, Social Security numbers, signatures, and identity-verification images
• Bank, card, credit, insurance, or investment statements and full transaction histories
• Private messages, intimate conversations, another person’s secrets, and information about children
• Confidential work, customer, employee, legal, security, and unreleased business material
• Whole contracts or records when the private terms themselves are what you want interpreted

For those tasks, use AI around the sensitive material instead of on it. Ask for a blank checklist, a list of questions for the responsible professional, a public explanation of unfamiliar terminology, or a template you fill out offline. Keep the real record beside you and outside the chat.

Settings do not erase the decision

AI services have different products, retention periods, training controls, business terms, and deletion behavior, and those can change. Check the official policy and settings for the exact service and account you are using. Stronger controls may matter to an organization with a reviewed agreement, but they do not turn an unnecessary upload into a good idea. If you cannot explain where the input goes and why the task needs it, keep it out.

The magic of AI is not worth pretending every problem belongs in a chat. Knowing when the input should stay with you is part of knowing how to use the tool.

• Share only: No personal document is needed for this framework; use the hypothetical examples and categories without testing them on your own sensitive material.
• Verify: A service’s current retention, training, deletion, sharing, and business-account terms in its official settings and policy, while treating uncertainty as a reason not to upload.

This is how to use AI.