Cloud Services
Shadow AI and the Privacy Act: What Changes on 10 December 2026
-
Dayle Wilson
- 7 min read
Summary
- From 10 December 2026, Australian Privacy Principles (APP) entities must explain in their privacy policy how computer programs use personal information to make or inform significant decisions about people.
- The OAIC reads “computer program” broadly, from spreadsheets to generative AI. A human signing off doesn’t automatically take you out of scope.
- Accurate disclosure depends on knowing which tools shape decisions. Shadow AI makes that question hard to answer.
- Start now: Inventory AI use, map it to decisions, and decide where sensitive AI should run.
Why your privacy policy is about to get harder to write
Ask most leadership teams which AI tools their people used last week. You’ll usually get a confident answer. The network traffic often tells a different story.
Alongside the approved tools, staff have found their own. A free chatbot drafts a customer email. A browser plug-in summarises a stack of CVs. None of it is malicious, but much of it is invisible.
From 10 December 2026, that invisibility becomes a compliance issue. Many Australian organisations must update their privacy policies to explain how computer programs shape decisions about people.
What changes on 10 December 2026
The Privacy and Other Legislation Amendment Act 2024 adds a transparency obligation to Australian Privacy Principle 1 (APP 1). The OAIC released its final guidance on 30 September 2026.¹ Under APP 1.7, the obligation applies when all three of these conditions are met:¹
- A program is involved. Your organisation has arranged for a computer program to make a decision, or to do something substantially and directly related to making one.
- The decision is significant. It could reasonably be expected to significantly affect an individual’s rights or interests.
- Personal information is used. The program uses personal information about that individual.
If you meet the test, APP 1.8 requires your privacy policy to describe three things:¹
- Personal information: the kinds your computer programs use.
- Automated decisions: the kinds those programs make on their own.
- Assisted decisions: the kinds where a program does something substantially and directly related to the outcome.
Notice the word “kinds”. You don’t need to publish source code or list every vendor. Still, the description must be meaningful to a reasonable person, and sensitive data such as health information should be clearly flagged.²
A human in the loop isn’t a free pass, either. A decision can fall in scope even when a person reviews the program’s output.² The OAIC considers machine learning or generative AI outputs used in significant decisions to be generally in scope. The exception is where extensive human oversight and control applies.²
Who's in scope, and what counts
The obligation applies to APP entities. Generally, that means businesses and not-for-profits with annual turnover above $3 million, plus Australian Government agencies. Some smaller organisations, such as health service providers, are also covered.
The OAIC’s definition of a computer program is broad. It covers rule-based processes, machine learning, everyday software and generative AI, including chatbots.² One of its own examples is simply a spreadsheet formula that ranks aged care clients by need.² Its non-exhaustive list of in-scope decisions includes:²
- Recruitment: software that sorts candidate profiles.
- Performance and pay: AI reports that rank staff or set bonuses.
- Lending: approving or rejecting loan and credit applications.
- Insurance: assessing eligibility for cover.
- Pricing: personalised pricing for significant goods.
- Health and disability: prioritising access to services.
The OAIC also walks through a manager using a generative AI tool to suggest staff bonuses. HR has oversight and a director approves every outcome. Even so, the tool remains in scope because managers rely on its output as a key factor.²
Enforcement is already under way. In January 2026, the OAIC began its first-ever privacy policy compliance sweep, covering around 60 entities across six sectors.³ Non-compliant policies can now attract infringement notices and penalties of up to $66,000.³
Why shadow AI complicates disclosure
Every privacy policy update starts with an inventory. For the AI your organisation bought and approved, that’s achievable. The challenge is everything else.
In a 2026 PagerDuty survey of 250 Australian office professionals, 70% had used AI at work despite believing policy didn’t allow it.⁴ Meanwhile, 40% had entered customer data into public AI tools.⁴
Two findings stand out for December. Some 39% had used AI for work advice or help with decision-making.⁴ A further 34% had shared AI-assisted work without saying AI was involved.⁴
Picture a hiring manager pasting shortlisted CVs into a free chatbot to rank them. They believe they’re simply working faster.
The OAIC hasn’t said every unsanctioned tool automatically counts as “arranged for”. However, it does say that relying on a program’s advisory output can bring a decision into scope.² It also urges organisations in doubt to take a cautious approach and disclose.²
That leaves two uncomfortable options. Disclose too little and your policy is inaccurate. Disclose broadly without evidence and you can’t explain it when someone asks.
Blocking popular AI tools rarely helps. It tends to move the behaviour rather than stop it. The same research found 44% of respondents had used personal devices for AI-related work.⁴ Once data leaves managed devices, you lose sight of it entirely.
How to close the visibility gap
No single product makes you compliant. In practice, organisations that handle this well focus on three things.
- Visibility: Base your inventory on evidence, not a staff survey. Identity logs, secure web gateways and cloud app discovery tools can show which AI services people actually use. Many organisations already own some of this capability through existing licensing.
- Control: General productivity AI and decision-making AI carry different risks. For significant decisions, know what data goes in and where it travels. Approved enterprise tools with data loss prevention help, as do private or sovereign environments you control.
- Oversight: In the OAIC’s bonus example, staff could interrogate outputs, narrow the tool’s parameters and document why they departed from its recommendation. That could indicate the output wasn’t substantially related to the decision.² It’s good decision hygiene, not a loophole.
What to do now
- Discover what’s really in use. Start with network and identity data, because people under-report tools they suspect aren’t allowed.
- List decisions that affect people. Hiring, performance, pricing, credit, eligibility and service priority are the usual suspects.
- Map programs to decisions. Include approved SaaS and vendor platforms, not just in-house tools.
- Group the personal information involved. Use kinds such as identity, financial or employment data, and flag anything sensitive.
- Ask your vendors. The OAIC expects software providers to explain, at a high level, how their tools make decisions.²
- Update and maintain your policy. Write in plain English, then revisit it whenever you adopt a new tool.
How Macquarie Cloud Services can help
There’s no single fix for December, but visibility and control are where most organisations start. On the visibility side, Microsoft Global Secure Access shows which AI tools people use, including unsanctioned apps. It can also stop sensitive data leaving through AI prompts, uploads and connectors, wherever people work. We run it as a managed service, backed by Managed XDR and our sovereign Australian SOC. That gives you evidence for your privacy policy and board, not just assurances.
Control is the other half of the picture. When AI shapes decisions about people, you need to know which data feeds it and where that data travels. Launch AI gives those workloads a private home in our sovereign Australian data centres, so the data flows you disclose match the ones that exist. If you’d like to see what AI is running in your environment today, get in touch with our team.
Frequently asked questions
Do I have to name the specific AI tools I use?
Not necessarily. APP 1.8 asks for the kinds of personal information and decisions involved.¹ Even so, you can’t describe those accurately without knowing which tools are in play.
Does this apply to small businesses?
Generally only if they’re APP entities. Most businesses under $3 million turnover are exempt, but some, such as health service providers, are covered.
Does a human signing off take us out of scope?
Not automatically. The OAIC says a decision can be in scope even when a person reviews the program’s output.²
What if we’re unsure whether a tool is in scope?
The OAIC recommends a cautious approach. If in doubt, include it in your privacy policy.²
The bottom line: From 10 December, your privacy policy needs to reflect how programs actually shape decisions about people. Shadow AI doesn’t read privacy policies, but regulators do. Get the inventory right, and the disclosure becomes far easier to write.
This article provides general information only and isn’t legal advice. Speak with your legal or privacy adviser about your specific obligations.
Resources
¹ OAIC, “New resources on transparency for use of AI and automated decision-making,” 30 September 2026.
² OAIC, “APP 1.7–1.9 Transparency Obligation Fact Sheet,” September 2026.
³ OAIC, “Privacy compliance sweep to put privacy policies under the spotlight,” 9 December 2025.
⁴ IT Brief Australia, “Australian office staff flout AI rules, study finds,” PagerDuty research, 6 July 2026.
Contact us
Talk to an expert.
Let’s talk about where you are and what better looks like for your business.
- 1800 004 943
- Level 15, 2 Market Street Sydney, NSW, Australia