Privacy Act 1988 for AI-Curious Operators: A Plain-English Guide

The Privacy Act 1988 isn't a wall between you and AI. It's a fence with a gate. This is a plain-English guide for the operations leader, founder or COO who needs to understand the rules well enough to ask their vendor the right questions — and reject the wrong answers — without becoming a privacy lawyer in the process.

Why this matters more in 2026

Three things have happened in the last two years that make AI and the Privacy Act collide more often:

1. AI is now embedded in everyday tools. Your CRM has AI. Your helpdesk has AI. Your email client has AI. Each of those processes data that, two years ago, lived only in your systems.
2. The OAIC is more active. The Office of the Australian Information Commissioner has been clearer and more public about its expectations around AI, automation and consent in 2025–2026.
3. Reforms are coming. The 2024–2026 reform tranches strengthen consent requirements, introduce a statutory tort for serious invasions of privacy, and clarify obligations around automated decision-making. The direction is more accountability, not less.

The thirteen rules in a paragraph each

The Australian Privacy Principles (APPs) are the backbone. Here's what each one means for an AI deployment:

APP 1 — Open and transparent management of personal information

You need a privacy policy. It needs to mention how you use AI (in plain enough English that someone can actually understand it). "We may use third-party AI to process data" is enough — but write it.

APP 2 — Anonymity and pseudonymity

Where practical, give people the option to deal with you anonymously. Most B2B operations are exempt in practice, but the principle matters for any consumer-facing AI chat or voice agent.

APP 3 — Collection of solicited personal information

Only collect what you need. AI agents have a habit of asking for everything ("can I have your name, email, phone, postcode, date of birth..."). Train them not to.

APP 4 — Dealing with unsolicited personal information

If your AI agent receives information it didn't ask for and doesn't need, you need a policy for what to do with it. Usually: destroy or de-identify.

APP 5 — Notification of the collection of personal information

When you collect, tell people what you're collecting, why, who you might share it with, and how they can access or correct it. AI tools must respect this — your privacy notice still applies when the chatbot is the one collecting.

APP 6 — Use or disclosure of personal information

You can only use information for the purpose you collected it (or a directly related secondary purpose someone would reasonably expect). Don't take support-ticket data and feed it into a sales-prospecting AI without thinking carefully about whether the customer would expect that.

APP 7 — Direct marketing

If you use AI to power direct marketing, the existing rules still apply — opt-in or opt-out, easy unsubscribe, real human accountability. AI doesn't get a pass.

APP 8 — Cross-border disclosure of personal information

This is the big one for AI. When your data goes to OpenAI's servers, Anthropic's servers, or Google's, it's crossing borders. You need to take reasonable steps to ensure the overseas recipient handles it consistent with the APPs. The practical mitigation: use enterprise tiers with contractual data protection, ideally with AU or equivalent data residency.

APP 9 — Adoption, use or disclosure of government related identifiers

Don't use Medicare numbers, TFNs, or other government identifiers for purposes other than what they were issued for. AI doesn't change this.

APP 10 — Quality of personal information

Information you hold must be accurate, complete and up to date. AI-generated summaries of customer information must be reviewable and correctable. Don't let the AI's interpretation become the record without a human in the loop.

APP 11 — Security of personal information

Reasonable security to protect against misuse, interference, loss, unauthorised access. The OAIC's expectation is that you'll have technical and organisational measures appropriate to the sensitivity of the data. For AI: enterprise APIs, encryption in transit and at rest, role-based access, audit logs.

APP 12 — Access to personal information

Individuals can request access to information you hold about them. That includes anything generated about them by AI (transcripts, summaries, scoring). Have a process.

APP 13 — Correction of personal information

If information is inaccurate, correct it. If your AI made an inaccurate summary about a customer and it's now in your CRM, you owe the correction the same as any other data error.

The four common AI mistakes that breach the APPs

1. Using consumer-tier AI for customer data

Pasting customer information into ChatGPT, Claude or Gemini's consumer products can mean that data is used to train future models. That's almost certainly a disclosure under APP 6 you didn't authorise. The fix: enterprise tiers only for anything touching customer data.

2. Treating AI vendors as one big anonymous third party

"We use AI" isn't enough on a privacy notice if customers are reasonably going to be surprised by which AI, where it processes, and what it does. Be specific enough that a curious reader understands the rough shape of it.

3. Storing AI conversation transcripts forever

Voice agent calls and chat transcripts are personal information when they include identifiable details. They need a retention policy aligned to your overall data retention. "Indefinite" is rarely justifiable.

4. Letting AI make decisions without human review on consequential matters

Refusing service, declining credit, terminating an account — anything that materially affects someone — needs a human in the loop. The reform tranches are explicit about this, and the OAIC's expectation is unambiguous.

A vendor selection checklist (use this on every AI tool)

1. Are you using an enterprise tier (not consumer)?
2. Does the vendor contractually commit to not training on your data?
3. Where does the data physically live? AU regions available where required?
4. Does the vendor have a current SOC 2 / ISO 27001 / equivalent?
5. Does the contract include breach-notification commitments compatible with your NDB obligations?
6. Can you delete data on request?
7. Is the data encrypted in transit and at rest?
8. Who, on the vendor side, can access your data — and when?
9. How is access audited?
10. What happens to your data if you cancel?

Any "no" or "we'll get back to you" on these is information. Often it's a reason not to proceed.

On-prem vs cloud — when each makes sense

Cloud (default for almost everyone)

Enterprise-tier APIs with AU residency and contractual no-training commitments cover the vast majority of business AI needs while remaining APP-compliant. This is what we use for ~90% of client builds.

On-prem (when you really need it)

For data that genuinely cannot leave your network — health records, legal privilege, certain government work — open-source models (Llama, Mistral, Qwen) run on your own infrastructure are the answer. Cost is higher (you're paying for GPUs and operations), but the data never leaves.

Hybrid (best of both)

Sensitive processing happens on-prem; non-sensitive enrichment, summarisation or formatting happens in the cloud with de-identified data. Architecturally more work but increasingly the right pattern for regulated industries.

What the 2026 reforms are likely to add

As of mid-2026, the second tranche of reforms is progressing. Key items for AI deployers to watch:

• Automated decision-making disclosure. A requirement to disclose when a decision affecting an individual is made (substantially) by automation. Plan to be transparent.
• Statutory tort. A new cause of action for serious invasions of privacy — including by AI systems. Increases the cost of getting it wrong.
• Stricter consent. Tightening around what counts as valid consent — opt-in becoming the norm for sensitive uses.
• Children's privacy. Specific protections that AI-powered consumer apps will need to plan around.

A five-step compliance checklist for your AI deployments

1. Inventory. List every AI tool in your stack and the data it touches.
2. Tier check. Confirm every one is on an enterprise tier with no-training commitment.
3. Disclosure. Update your privacy notice in plain English. Don't hide it.
4. Data flow diagram. Sketch where data physically goes. This is also useful for the NDB scheme.
5. Human-in-the-loop on consequential decisions. Define what's consequential and require human sign-off on those decisions.

This is not legal advice. For anything material, talk to a privacy lawyer. But you'll be a much better client if you arrive with this groundwork done.