What staff should never paste into an AI tool

Keep three things out of an AI chat box: anything that identifies a person, anything the business promised to keep private, and any password or key. When in doubt, leave it out.

Why a chat box isn't a private notebook

Australia's privacy regulator recommends, as best practice, keeping personal details out of public AI chatbots, and sensitive details most of all.¹ Its guidance describes staff at an insurance company who try a public chatbot on a customer's claim, health details included.¹ The guidance says that by doing so, the company is disclosing the information to the owners of the chatbot.¹

That advice is written for Australian businesses. Elsewhere the rules differ, so read your own regulator's page.

What counts as personal information

More than names. The regulator says a detail can count as personal information even when it's false.¹ So an AI tool's wrong guess about a named person can still count. Sensitive information is a narrower group within it, and it generally gets stronger protection.¹

A practical "never paste" list

The regulator recommends not entering personal or sensitive information into publicly available chatbots.¹ The "never" in this table is a suggested house rule, not the regulator's wording.

Why secrets are a special worry

AI tools can be tricked. The UK's cyber security centre calls prompt injection one of the most widely reported weaknesses in large language models, and says it can make a model reveal confidential information.² Passwords are the clearest case. No task needs one pasted in. If a staff member has already pasted one, our suggestion is to change it straight away and tell the owner. This is our own suggestion, not from a source.

Getting help without the real details

Staff can still get value from a tool. The trick is to give it the shape of the job, not the people in it.

• Replace names with labels. "A customer has complained that a delivery was late. Draft a calm reply."
• Remove identifiers first. Strip out account numbers, addresses and dates of birth before pasting a block of text.
• Describe, don't attach. Explain what a contract clause does instead of uploading the contract.
• Use sample data. For spreadsheets, build a fake table with the same columns.

The regulator expects organisations to assess whether there are ways to minimise the amount of personal information that goes into an AI product.¹ Taking out names is one such way, but it isn't foolproof. A detailed story about a small town's only bakery can still point to one person.

Writing it into the policy

The regulator lists, among measures a business might consider, prohibiting staff from entering personal information in AI inputs. It adds that any such controls or prohibitions need to come with training and auditing.¹ A short rule might read:

"Don't enter customer or staff personal details, passwords, or confidential documents into an AI tool. The only exception is a use the approved-tools list allows. Unsure? Ask [name]."

Add three real examples from your own business. People remember a real case better than a label like "client data".

Our own suggestion, not from any source: name who staff should tell the same day if they paste something they shouldn't have. For what to do about leaked personal information, use your own regulator's pages on data breaches.

For related rules, see the approved tools list and checking output.