AI on Australian Government Websites: Data Sovereignty Requirements

Sasha Shevelev
Sasha Shevelev 17 November 2026

Put a chatbot on a government website and sooner or later a resident will type something personal into it. A Medicare number. A diagnosis. A long complaint about the agency itself. Nobody asked them to. It doesn't matter: from that moment, every rule about where that information may go applies.

Agencies often ask the right first question ("is it onshore?") and then stop. Onshore is necessary for most of them. It isn't sufficient.

Several instruments apply at once, they bind different public bodies, and some are law while others are policy. Here's what each demands of an AI feature on a public website, and what to get from a partner in writing. It reflects October 2026 and isn't legal advice. Nobody, us included, has every edge of this worked out yet.

Key Takeaways

  • Commonwealth agencies, state agencies and councils each answer to different rules.
  • An Australian region doesn't settle APP 8 or foreign jurisdiction by itself.
  • The OAIC expects public-facing chatbots to be clearly identified as AI.
  • From 10 December 2026, APP entities' privacy policies must cover software that uses personal information to make or substantially inform significant decisions about people.
  • Can't say where the transcripts end up? Then you aren't ready to launch.

Why "Is It Onshore?" Became the First Question

Government sites are adding AI fast: chat that answers residents, search that understands intent, help with long forms. Each routes citizen input through a model plus logging, analytics and an index of your pages. Residency (where the bytes sit) and jurisdiction (whose law reaches them) aren't the same thing, and the instruments below care about both.

The Privacy Act 1988 and the APPs

The Australian Privacy Principles apply to most Australian Government agencies. State agencies and councils generally fall under their own state's law instead, which in NSW means the Privacy and Personal Information Protection Act 1998 (the PPIP Act). For a Commonwealth agency, four principles do most of the work.

  • APP 1, transparency. A clearly expressed, up to date privacy policy, including whether personal information is likely to go overseas. From 10 December 2026, APP 1.7 to 1.9 (added by the Privacy and Other Legislation Amendment Act 2024) also require it to cover computer programs that make, or do something substantially and directly related to making, decisions that could significantly affect someone's rights or interests, where their personal information is used.
  • APP 6, purpose. Information collected for one purpose can't be used for another unless an exception applies. Feeding resident questions into model improvement is the obvious trap.
  • APP 8, cross-border disclosure. Before disclosing personal information overseas, take reasonable steps to ensure the recipient doesn't breach the APPs. Under section 16C you're generally accountable if it does.
  • APP 11, security. Reasonable steps against misuse, interference, loss and unauthorised access, modification or disclosure. Chat transcripts count.

One nuance catches people. The OAIC says giving personal information to an overseas recipient, including a contractor such as a cloud provider, is in most circumstances a disclosure. Only in relatively limited circumstances, where the agency keeps effective control, might it be a "use", and even then the agency may still be held accountable. (So "it's only storage" is weaker than it sounds in a sales meeting.)

What the OAIC's AI guidance adds

The OAIC's guidance on privacy and the use of commercially available AI products (October 2024, updated January 2025) is written for organisations deploying AI, but it's a sound benchmark for agencies too. On a website, the points that bite:

  • Obligations apply to personal information put into an AI system and to output containing it. AI-generated or inferred personal information is a collection.
  • Public-facing tools such as chatbots should be clearly identified as AI, and privacy policies should explain the AI use.
  • Some products let the vendor collect customer input for further training. Check the terms and settings.
  • Take a privacy-by-design approach, including a privacy impact assessment.

The DTA Policy: Governance, Not Hosting

The Policy for the responsible use of AI in government is policy, not legislation, though it's mandatory for those it covers. Version 2.0 took effect on 15 December 2025 for non-corporate Commonwealth entities, with carve-outs for defence and the national intelligence community, and minor updates followed as version 2.1 on 21 October 2026. It requires accountable officials, an owner for each in-scope use case, an internal use case register shared with the DTA, a published AI transparency statement reviewed at least annually, and impact assessments for in-scope use cases.

Is a resident chatbot in scope? Quite possibly; the agency's assessment decides. The policy is about accountability; hosting is settled elsewhere.

PSPF, the ISM, IRAP and the Hosting Certification Framework

This is the layer that governs where government data may live, and the one most often blurred in proposals. If four acronyms in a row make your eyes glaze, fair enough. They weren't written as a set.

  • The PSPF, administered by the Department of Home Affairs, sets mandatory protective security requirements for non-corporate Commonwealth entities. Release 2026 was issued on 1 July 2026.
  • The ISM is ASD's cyber security framework, applied through each organisation's own risk management.
  • IRAP assessors are endorsed by ASD to assess systems and cloud services against the ISM. ASD notes an assessment doesn't by itself mean a system complies with the controls tested, and the agency still makes its own authorisation decision. (ASD stopped certifying cloud services in 2020.)
  • The Hosting Certification Framework certifies data centre and cloud providers. Run by Home Affairs since 1 May 2023, it requires sensitive government data, whole-of-government systems and PROTECTED systems to be hosted on certified services. It's mid-reform, with new registrations paused since 3 November 2025, so treat confident claims about its final shape with suspicion.

The buyer's question is short. What classification of data will this feature touch, and are we Commonwealth, state or local? And remember the HCF certifies the provider, not your application. A certified host doesn't make the chatbot compliant.

The CLOUD Act, and Why a Region Doesn't Settle It

The US CLOUD Act of 2018 requires US providers to produce data in their possession, custody or control under US legal process, whether it's stored inside the United States or not. A US-headquartered provider running an Australian region is still a US provider.

We offer AI integration hosted in Australian Azure regions, so it'd be convenient to tell you a local region settles everything. It doesn't. But it's a factor to weigh, not a reason to avoid the big clouds, whose Australian regions bring mature controls and published assurance material. Choose deliberately and record why in the privacy impact assessment. I'd rather an agency write down "we accepted this residual risk, for these reasons" than pretend the question never came up.

The State Layer: NSW as the Worked Example

NSW changed its rules this year. The NSW AI Operational Policy, published on 30 July 2026 and made mandatory for NSW Government agencies by Circular DCS-2026-02, requires an accountable official, registration of every AI use case on the AI Assessment Framework platform (mandatory for new registrations from 30 September 2026), the framework itself where registration calls for it, and referral of high and critical risk use cases to the AI Review Committee.

Councils aren't bound by that policy. They are bound by the PPIP Act, including the Mandatory Notification of Data Breach scheme in force since 28 November 2023, so a council chatbot that exposes transcripts can still become a notifiable breach. Other states differ; check yours.

What This Means for the Website Itself

For a chatbot answering residents or AI-assisted site search, trace each of these to a region and a legal entity:

  • Prompts, in transit and at the model endpoint.
  • Logs and transcripts, which several services may each keep, on different retention clocks.
  • Embeddings built from your content. Confirm nothing internal crept in.
  • Monitoring. Ask whether the model provider retains inputs for abuse review, and for how long.

Is resident input used to train anything? (APP 6 territory.) Do the privacy notice and the chat window both say AI is answering? Who reads transcripts, and when are they deleted? If the assistant pre-fills an eligibility form, is it edging towards a decision under APP 1.7?

Most of this is AI integration work, not model work. Grounding answers in the agency's published content, rather than training on what residents type, keeps these questions answerable.

What to Demand From a Partner, in Writing

Verbal assurances evaporate at audit time.

  1. Hosting region and tenancy. Good answer: a named region and whose tenancy. Weak answer: "it's all in the cloud, fully secure."
  2. Every subprocessor and model endpoint, and where each runs.
  3. Retention and training. "We don't train on your data" should arrive with the clause or setting that makes it true.
  4. Audit trail and export. Can you get transcripts out in a usable format?
  5. APP 8 accountability. If anything goes offshore, which terms cover it?
  6. Precise assurance claims. If someone cites IRAP or the HCF, ask for scope and date. Be wary of anyone, web agencies included, claiming to be "IRAP certified". That isn't what IRAP does.
  7. Exit. How data is returned and deleted when the contract ends.

When the Answer Is "Not Yet"

Sometimes the sovereign-safe answer is to hold off. If the data is classified above what the hosting arrangement supports, wait. If nobody can say where transcripts go, wait. A good FAQ and decent conventional search serve residents fine meanwhile. I'd rather an agency ship that than an assistant it can't account for.

Where Webcoda Fits

Webcoda is a Sydney web development agency that builds government websites and intranets and offers AI integration with data kept onshore in Australian Azure regions or the client's own tenancy, confirmed in writing. That's a hosting offer, not an accreditation, so the assurance questions above still belong with the hosting provider and your own security team.

Based in Surry Hills and working since 2005, Webcoda is B Corp certified and a Microsoft Partner. Government clients include the Australian Federal Government, NSW Health, the AIHW, NESA and Newcastle City Council, and published work includes intranets for NSW Justice Health, NSW Health 1 Reserve Road and Mid North Coast LHD, plus the Infrastructure NSW website. There's more on the government sector page, and if the real question is whether you're ready at all, start with an AI readiness assessment.

Frequently Asked Questions

There's no single blanket law saying so, but in practice it usually should be. For Commonwealth agencies, APP 8 and section 16C of the Privacy Act keep you accountable for overseas recipients, and the PSPF and Hosting Certification Framework tighten as data classification rises. Onshore hosting simplifies all of that, though it doesn't settle foreign jurisdiction questions such as the US CLOUD Act on its own.

Generally not. The Australian Privacy Principles apply to most Australian Government agencies, while state agencies and councils fall under state privacy law. In NSW that's the PPIP Act, which since 28 November 2023 has included a Mandatory Notification of Data Breach scheme covering councils as well as state agencies.

It's policy rather than legislation, but it's mandatory for non-corporate Commonwealth entities. Version 2.0, effective 15 December 2025 and updated to version 2.1 on 21 October 2026, requires accountable officials, a published AI transparency statement, an internal use case register and impact assessments for in-scope use cases. State agencies and councils aren't covered; NSW agencies have their own AI Operational Policy.

No. IRAP is an assessment by an ASD-endorsed assessor against the ISM, not a certification, and each agency makes its own authorisation decision. Certification of hosting providers is a separate scheme, the Hosting Certification Framework, now run by Home Affairs, which paused new registrations from 3 November 2025 while it's reformed.

It should explain how AI is used and how personal information entered into it is handled, and the chatbot itself should be clearly identified as AI, per the OAIC's guidance. From 10 December 2026, APP entities must also describe computer programs that use personal information to make or substantially inform decisions significantly affecting people's rights or interests.

Talk to Webcoda

Webcoda is a Sydney-based, B Corp certified digital agency that has been building websites and enterprise web applications since 2005, for organisations including Sony, BridgeClimb, the Australian Federal Government and NSW Health. We are an Umbraco Gold Partner, a Gold partner for Xperience by Kentico and a HubSpot Solutions Partner, so platform advice comes without a single-vendor agenda.

Get in touch to talk through your project, or browse our full range of services and platforms.