Consent Won't Cover Your AI: Australia's New Privacy Test

For as long as Australian firms have handled client data, the privacy defence has been simple: the client consented, so the data was fair to use. The exposure draft the Attorney-General released on 31 August 2026 retires that logic. Its centrepiece is a "fair and reasonable" test — you may collect, use or disclose someone's personal information only where doing so is fair and reasonable in the circumstances, judged objectively. Consent no longer buys your way past it. Submissions on the draft close today, 18 September.

For a firm running client data through AI, a second line matters just as much. The same draft widens the meaning of "collection" to cover information your systems generate, derive or infer — including inferences drawn by artificial intelligence. In plain terms: your model's guesses about a client are now regulated data.

On 31 August, Attorney-General Michelle Rowland released the exposure draft of the Privacy Amendment (Personal Data Protection) Bill 2026 — the long-awaited second tranche of Privacy Act reform, some 40 proposals in all. The one that resets the ground rules is a new fairness standard: every act of collecting, using or disclosing personal information must be fair and reasonable in the circumstances, weighed against what a reasonable person would expect, whether less data would achieve the same purpose, and whether the individual had a genuine choice. As The Conversation put it, a privacy policy buried in legal jargon "will not save a practice that an ordinary person would never expect". It is, on the government's account, a world-first — a step past the notice-and-consent model the US relies on and the legal-basis model of Europe's GDPR.

One thing the draft leaves out is telling. The small-business exemption — the $3 million turnover line that keeps most small firms outside the Privacy Act entirely — is untouched. The government has agreed in principle to remove it, but not in this bill. So the fairness test bites first on the firms already inside the Act's perimeter.

Want to know what this means for your business? We'll map it in one call.

Two groups of our readers are squarely in scope. Any firm over the $3 million line is one. The other is larger than it looks: the accounting, legal, conveyancing and real-estate practices that came under the Privacy Act's reach this year through the AML/CTF reforms, plus health-service providers — GPs, physios, psychologists, allied health — who are covered regardless of turnover. If you handle sensitive client information and you have started pointing AI at it, the fairness test is your problem now, not a future one.

Here is where it gets concrete. Feed a client's file into a US-hosted AI model to summarise it, and the question is no longer "did they consent" but "would a reasonable client expect their financial records to be processed offshore, and could you have done it with less of their data". Use AI to score clients for risk or likelihood to pay, and every inference the model draws — a health flag guessed from a gap, a creditworthiness signal — is now "collected" personal information subject to the same standard. The inference the client can't see and would never expect is exactly what the test is built to catch. This is a Customer Intelligence capability with a compliance line drawn through it. Map where AI already touches client data in your firm before the standard is settled, not after.

None of this displaces the obligation already locked in: from 10 December, every automated decision that significantly affects a client must be disclosed in your privacy policy. The fairness test is the standard; the December rule is the disclosure. Firms in scope now carry both.

Don't wait for the bill to pass — the practices it targets are ones you can find today. Three moves. First, inventory the personal information flowing into AI: which tools, whose data, and what each one infers or derives, not just what you type in. Second, run each use through the fairness question instead of the consent one — would a reasonable client expect this, and could you achieve the purpose with less of their information? A defensible answer is about the work; "they ticked the box" is not. Third, retire the uses that fail — the ones leaning on a buried policy line — and make sure a human can explain any AI-generated inference that affects a client's rights or money. If you're under the $3 million line and still exempt, do it anyway: it's the standard heading your way, and clients have started asking who touches their data before any regulator does.

Key takeaways

Common questions

Related notes

Is your AI use fair and reasonable — or just consented to?

Most firms can't yet say what their AI tools infer about clients, or whether a reasonable client would expect it. That's the gap the new test is built to expose, and it's a short conversation to close: which tools touch client data, what they derive from it, and which uses won't survive a fairness test. Book a call and we'll map it against the draft while you still have runway.