Privacy when customer feedback goes into an AI tool

Don't paste customer comments into an AI tool until you've asked three things. What personal details are in them? Where does the tool send and keep them? And does your country's privacy law apply to you?

None of this is legal advice. Privacy rules differ from country to country, and sometimes by state or industry. For a ruling on your own business, ask your national privacy regulator or a lawyer.

What counts as personal information

It's wider than a name. The Australian regulator says personal information covers a broad range of information, or an opinion, that could identify a person.¹ That is the Australian wording. Other countries use their own terms and their own definitions.

In feedback, personal details turn up in odd places:

• The name signed at the bottom of a comment.
• An email address or phone number typed into a free-text box.
• "My daughter Emma, who's in your Tuesday class".
• A detail that points to one person, such as "the tall man on the late shift at your Hobart shop".
• Health, money or family details people add on their own.

Nobody asked for these. Customers type them anyway. A free-text box has no filter, and that's the whole problem.

What an Australian regulator advises

Australia's privacy regulator, the OAIC, publishes guidance on commercial AI products. Its advice, as best practice, is to keep personal information out of publicly available generative AI tools. Sensitive information gets a special mention.²

That advice is from Australia, and it speaks to organisations covered by the Australian Privacy Act. Elsewhere, look up your own regulator's view. In the UK that's the Information Commissioner's Office. In the US, start with the Federal Trade Commission.

The same guide lists questions to ask about a product. One is who will be able to get at the personal information that goes in or comes out.²

Take out what the tool doesn't need

The tool doesn't need to know who said what. It only needs the words. A privacy idea called data minimisation fits here. It means gathering and keeping only the personal information a specific purpose needs.³

So before you paste:

1. Delete the name, email and phone columns from the copy you'll use.
2. Search the comments for @ signs, long digit runs and common first names.
3. Replace details that point to a person with a plain label, such as NAME or SHOP.
4. Keep the full original file somewhere separate, with its link to each customer.
5. Give each comment a number, so you can match results back to the original.

Searching misses things. Always read a sample of the cleaned file yourself before you send any of it anywhere.

Questions to ask about the tool

The answers sit in the maker's own terms and privacy pages. Terms change, so write down the date you read them.

• Does it keep what you paste, and for how long?
• Can the maker read it, and can staff review it?
• Is it used to train the maker's models?
• Is there a business or paid version with different data terms from the free one?
• Where is the data stored, and which country's law covers it?
• Can you delete a conversation, and does deleting it remove it for good?

The OAIC adds that reviews of the product, staff training and monitoring should carry on while it's in use. Treat this as a check you repeat.²

Customers and what they were told

Look at what your survey form or privacy notice told people about how their answers would be used. Pasting their words into a new tool may be a different use from the one you described.

Still unsure? Put the question to your regulator or an adviser before you paste.

Accuracy is a privacy matter too

The same guide warns that AI output can be wrong.² A wrong summary about a named person can do harm. That's another reason to keep names out.

Checking output is covered on the checking page. For the wider view, start at the first lesson.