Before You Paste Work Into AI: 3 Checks to Share Less Private Data

You only need help with one small problem, but the file contains everything around it. How much does the AI really need to see?

The document is already open.

Maybe you want AI to make one email clearer. Perhaps one spreadsheet formula is broken. Maybe one part of your code gives an error you do not understand.

Copy. Switch window. Paste.

It can happen in seconds.

But your question may be only a tiny part of what you copied. Around it could be a customer’s name, a colleague’s phone number, an internal price, a private note or a secret access key that has nothing to do with the answer.

So before we ask, “Can AI solve this?”, let’s ask three smaller questions.

A safer three-question check

• Let’s ask: What does the AI actually need?
• Let’s ask: Can we make the example smaller?
• Let’s ask: Is this information suitable for this exact tool and account?

1. Ask what the AI actually needs

Imagine a customer sends a complaint and you want help making your reply clearer.

The original message contains a real name, email address, order number, delivery address and a long explanation of what happened.

But your real question is smaller:

“How can I make this reply calm, clear and easy to understand?”

The AI may need to understand what went wrong. It may not need to know who the customer is.

That difference matters.

In guidance for small and medium businesses, the Australian Cyber Security Centre says that uploading customer or staff details to generative AI tools — systems that create new content from instructions — without properly removing identifying information can expose private or sensitive information. The same guidance recommends removing or changing personal details so they cannot be linked to a person when those details are not needed. (Australian Cyber Security Centre)

So look once at what you are about to send and separate the problem from the people and records around the problem.

For example:

“Maria Lopez at maria@example.com says order 583104 has not arrived.”

could become:

“A customer says an order has not arrived.”

You still have the part needed to discuss the reply. The identifying details are gone.

The same check can catch phone numbers, account numbers, employee records, private business figures and secret access information. An access key, for example, is a secret code that software can use to connect to a service. If it is not needed to solve the question, it does not need to travel with the prompt.

The point is not “never give AI any information.”

It is simpler:

Do not confuse everything inside the file with everything needed to answer the question.

2. Ask whether a smaller example can keep the same problem

Removing names is a good start, but sometimes you can reduce the material much further.

Imagine a spreadsheet with thousands of rows.

One formula is returning the wrong result.

If the mistake depends only on three columns and the formula itself, the other rows may add information without helping to explain the error.

Instead, you might make a small example:

Item	Expected	Result
Sample A	20	20
Sample B	35	0
Sample C	15	15

Then you can show the formula that produces the wrong value.

The real customer list, prices and unrelated rows are no longer part of the question.

The Australian Cyber Security Centre makes a similar point in its example for secure AI chatbots: collect only the information that is essential for the chatbot’s function. That example is written for businesses building chatbots, but it gives us a useful practical question when we prepare our own prompt: what is essential for this task? (Australian Cyber Security Centre)

You can use that question in different kinds of work.

A long email chain can become a short summary.

A large spreadsheet can become a few made-up rows.

A whole program can become the small part where the error happens.

Real names can become “Person A” and “Person B.” Real amounts can become invented amounts when the exact values do not affect the problem.

This gives you a useful test:

If I replace the real details, does the same problem still happen?

If yes, the smaller example may be enough.

There is also a limit.

Sometimes the real details are the reason the problem exists. A spreadsheet error may appear only with a certain type of value. A program may fail only under a particular setup. A question may depend on the exact wording of a document.

In that case, removing too much can destroy the problem you are trying to show.

So the goal is not to make the prompt as empty as possible.

The goal is to find the smallest useful version: enough information to preserve the problem, but no unrelated information carried along with it.

3. Ask whether the real information belongs in this exact tool and account

Suppose you have reduced the material as far as you can, but some real information is still necessary.

Now the question changes.

It is no longer only:

“What should I remove?”

It becomes:

“What happens to information I send through this particular service?”

That cannot be answered by the word “AI” alone.

Different providers, products, account types and settings can handle submitted information differently. The Australian Cyber Security Centre therefore recommends reviewing an AI provider’s settings, terms and privacy information to understand how submitted data may be collected, stored and used. (Australian Cyber Security Centre)

OpenAI provides a current example of why the exact product and account matter.

As checked on August 28, 2026, OpenAI says that content from its services for individuals, such as ChatGPT and Codex, may be used to train its models — in simple terms, to help improve how future models behave. It says users can opt out so new conversations are not used for this training. OpenAI separately says that, by default, inputs and outputs from ChatGPT Business, ChatGPT Enterprise and its API Platform, which developers use to connect software to OpenAI models, are not used for model training unless an organisation explicitly opts in to data sharing. OpenAI also notes that some choices, such as voluntarily sending product feedback, can have separate data-use rules. (OpenAI Help Center)

Those statements describe OpenAI’s current services. They are not rules for every AI provider.

That is why “Does this company train on data?” can be too broad a question.

A better question is:

How does this exact product, account and setting handle what I am about to send?

Work and school can add another check. If your organisation has rules about which AI tools may receive internal information, those rules matter before you paste the material. A personal account may also have different controls from an organisation-managed account.

Imagine that your finger is above the paste shortcut and you cannot answer these questions:

What information am I sending?

Why is it necessary?

Do I know how this exact service and account handle it?

Am I allowed to send it here?

You do not need to become a cyber-security expert before using AI.

But if you cannot answer those questions, you can keep the real material out of the prompt while you check. You can remove more information, read the current settings and privacy information, or use the tool your organisation has approved.

The useful moment is the one before the information leaves your hands.

First find the need, then shrink the material, then check where it is going

The three checks solve different parts of the same problem.

The first separates your actual question from identifying or private details that happen to sit beside it.

The second reduces the remaining material to the smallest example that still contains the problem.

The third matters when real information still has to remain. It asks whether the exact service, account and rules make that information suitable to send.

The order matters.

Checking a tool’s privacy settings does not make unnecessary personal information necessary. Removing a person’s name does not make every remaining business record suitable to upload. And reducing a file does not help if the small piece you kept still contains information that should not go into that system.

So the goal is not simply “be careful with AI.”

It is a clearer way to think:

Need it. Shrink it. Check where it is going.

When those three questions become part of the moment before you paste, you can still use AI for useful work without automatically handing it the whole document.

Move closer to the person you want to become

• Let’s pause over the next real document for a moment and separate the question from the information surrounding it.

• Let’s check the exact service and account when real personal or work information still needs to remain.