What should I keep out of an AI assistant?

Three agencies, one answer

Three Australian authorities answer the question from three directions: the OAIC on privacy, the National AI Centre on adoption, and IP Australia on the business's own secrets. On what to keep out, they point the same way.

Read together, the three give one test: the kind of information decides the risk, not the polish of the tool. The OAIC states it as a rule with a best practice attached, so it applies to what goes in and to what comes back that contains personal information. The National AI Centre states it as a caution, aimed at what you would call sensitive or confidential, and at browser-based tools in particular. IP Australia states it as the business's own assets: the recipe, the list, the idea not yet filed.

What counts as personal information

The OAIC's guidance on training generative AI models gives the working definitions, and both are wide. Personal information includes a broad range of information, or an opinion, that could identify an individual: it may include information such as a person's name, contact details and images or videos where a person is identifiable.

Sensitive information is a subset of personal information that is generally afforded a higher level of privacy protection. The OAIC's definition reaches biometric information to be used for automated biometric verification or biometric identification, biometric templates, health information about an individual, genetic information, and personal information about certain topics such as racial or ethnic origin, political opinions or sexual orientation.

Neither definition asks whether the information is interesting, or whether anyone would mind. It asks whether a person could be identified from it, and for the sensitive list, whether the information is of a kind the Act singles out. A scribbled note about a customer's health sits in the narrower set; the customer's email address sits in the wider one.

The customer list: two kinds of risk at once

A customer list is the ordinary case that sits in both sets at once. IP Australia names customer lists first among the confidential information a business relies on. The contact details in the same list are personal information under the OAIC's definition, because they could identify an individual.

So one paste can put a business secret and personal information at risk in the same act. The two losses are answered in different places: the confidentiality of the list is the business's own asset, and the personal information in it carries the OAIC's best-practice line about publicly available generative AI tools.

Work you do not own

IP Australia's guide to using AI to create IP asks a business to consider whether anything it uploads into an AI tool via prompts, such as pictures, drawings or text, could be protected IP. Its warning is specific: in some cases, uploading works to an AI model that you do not own the rights for could infringe on the rights of other IP rightsholders.

This is the part of the In leg that has nothing to do with privacy. A picture, a drawing or a passage of text that belongs to someone else: the question is not whether it is sensitive, but whether the rights in it are yours to upload. IP Australia's guide puts the consideration before the upload, not after it.

What to do instead

The sources do not leave the question open. The OAIC's best practice is to keep personal information, and particularly sensitive information, out of publicly available generative AI tools. The National AI Centre's caution points the same way for sensitive or confidential information, with a particular eye on platform-based or browser-based tools.

Where a business decides it must ask an assistant about personal information, the Act's use and disclosure rule applies to what goes in. APP 6 provides that an individual's personal information can only be used or disclosed for the purpose or purposes for which it was collected, known as the primary purpose, or for a secondary purpose if an exception applies. If personal information is being input into an AI system, APP 6 requires entities to only use or disclose the information for the primary purpose for which it was collected, unless they have consent or can establish the secondary use would be reasonably expected by the individual, and is related, or directly related for sensitive information, to the primary purpose.

One more protection sits in how the purpose is read. How broadly a purpose can be described will depend on the circumstances, but in cases of ambiguity, the OAIC considers that the primary purpose for collection, use or disclosure should be construed narrowly rather than expansively.

The practical order that follows is short. Decide what is in the text before it goes: a name, a health note, someone else's work, a recipe. If any of those is in it, think again before it goes in. The OAIC's best practice keeps personal information out of publicly available generative AI tools, and APP 6 limits its use to the purpose it was collected for unless an exception applies. The OAIC's warning is that once personal information is in, it will be very difficult to track or control how it is used.