Skip to main content
The Human Bit

Lesson 04 · 8 minutes

Know what not to share

Sort your information before it goes near an AI product, and send the least you can while still getting a useful result.

Recognise sensitive inputs and choose a safer way to work.

Why this matters

Convenient features make it easy to hand over sensitive information without noticing you crossed a line.

Controls differ by product, by plan and by organisation, so a familiar brand name is not the reassurance it feels like.

The practical bit

Before you paste or upload, ask who the information identifies, who owns it, what policy or agreement covers it, and what harm could follow if it leaked. Then use the smallest amount that still lets you do the task.

Strip out names and any detail you do not need, stay inside the tools your organisation has approved, or build an invented example instead. If the privacy, retention or training terms are unclear, whether your data is stored and whether it is used to improve the AI, stop and confirm them rather than guessing.

Watch the word memory too. Products use it for at least three different things: the current chat, the wider context the model can see, and information saved for later, such as stored memories, past chats it can reference, indexed project files and connected apps. A single switch marked off may not cover all of them, so check what each setting actually turns off.

What it looks like

At work

Test a staff-letter workflow with invented names and situations first. Only once it works do you go near real employee information, and then only through your organisation's approved tools and policy.

In everyday life

Ask for help understanding how a medical form is structured without uploading the filled-in version, and treat the explanation as background rather than as medical advice.

Common mistake

Removing a person's name but leaving in the role, dates, location or case details that still point straight to them.

Use your judgement

Would you be comfortable emailing this exact information to an unknown external service?

Try it now

Minimise one input
  1. Choose a document or task you might give an AI product.
  2. Mark personal, confidential, commercially sensitive and unnecessary details.
  3. Create a reduced or synthetic version that can still test the workflow.
Check: The test should work without exposing information unrelated to the result.

Use this in a real situation

Practise with a permitted fileWork through a file example without handing over sensitive material.Open the practical →

Keep this

Use the least sensitive information that can still give you a meaningful test.