The Sovereignty Gap: Why Cheap AI Comes With Real Risks for Australia
I've been full-time in the AI space since last year, working with Australian tech teams. We've all seen the toggle: "Do not use my data for training." We click it, breathe a sigh of relief, and keep building.
But here's the uncomfortable truth: every major model provider still captures and retains request/response data for legal, safety, and compliance reasons. Anonymised datasets can still be re-identified through modern AI mosaic attacks. And in the process of learning patterns of malicious behaviour, edge-case IP and sensitive ideas can quietly migrate into internal models.
The Data We're Actually Sharing
On the ground, I'm seeing teams continue to paste information into these tools without really thinking through what tiers of data are being processed and logged. It sits dormant until, one day, it either gets deliberately used or mistakenly surfaces.
The financial math that moves entire economies is at play here: massive capital flowing into a handful of frontier players, uncertainty about who the winners and losers will be, questions around their ethics, and how much real control nations and societies actually retain while trying to keep an open marketplace alive.
Add in the job shifts already happening and the way other critical research domains are quietly starved of funding, and the stakes feel existential.
The Real Trade-Off
It feels uncomfortably familiar. The same dynamic we saw with earlier platforms where the service was free and the real product was the user. Only now the stakes are higher: our ideas, our research, our public-service data (health, aged care, education, pensions, disability support), and ultimately our national capability.
The trade-off is real. Frontier models through cloud APIs are dramatically cheaper and more capable than self-hosting or local rigs for the vast majority of use cases. Market forces and network effects are already steering us toward a handful of dominant providers. That's great for speed and iteration. It's risky for long-term control.
Especially when we reach AGI and the shortest pathway to prosperity for one business or nation might simply be: look at all the existing data and amplify your wealth and power, so we can do good for humanity.
What Australia Is Actually Doing
Australia is not standing still. The National AI Plan (December 2025) and the APS AI Plan, with its GovAI sovereign platform, show genuine intent: keeping public-sector data inside Australian infrastructure, establishing the AI Safety Institute (now operational), and focusing on practical productivity gains across health, education, aged care and more.
These are meaningful steps. The government is investing in sovereign capability, not just talking about it.
Yet for national security especially, the question is sharper: do we really want to depend on providers like Palantir or Anduril for the capabilities that protect us? Most Aussies, easy-going and with the right heart, would rather spend that money looking after our people. But the world is not a super-safe space.
You will need capabilities. Period. Or we put our next generations at the mercy of others and their shifting notions of righteousness.
The speed with which friends can become adversaries (US tariffs being a recent reminder) should give every policymaker pause.
What Engineers Are Actually Implementing
From the trenches, here's what practical controls look like today:
- Data tier classification: Teams are starting to classify data before it touches an LLM. Public, internal, confidential, restricted. Different tiers get different treatment.
- Request logging and audit trails: Every API call logged with metadata. Who sent what, when, to which provider. Retention policies that actually get enforced.
- Local model deployment for sensitive workloads: Running Llama or Mistral on-premises for anything that touches customer PII or proprietary IP. Yes, it's slower. Yes, it's more expensive. But you control the data.
- API gateway controls: Rate limiting, content filtering, and automated redaction before requests leave the network perimeter.
- Vendor contract scrutiny: Actually reading the data processing agreements. Understanding where data is stored, who has access, and what happens when you terminate the contract.
These aren't theoretical. These are the controls that separate teams who are thinking about sovereignty from teams who are just hoping for the best.
The Conversation We Need
This isn't about rejecting powerful AI. It's about ensuring we harness it on terms that reflect Australian values, starting at home, while still contributing responsibly to global progress.
We need genuine, cross-domain perspectives:
- AI engineers building production systems: What practical controls are you actually implementing today around data tiers and logging?
- Frontier thinkers and educators: How do we make sovereignty, accessibility, and open-market innovation compatible at scale?
- Global safety leaders and academics: Where should the balance between open innovation and protective guardrails sit, especially as AGI pathways emerge?
- National-policy, privacy, and security experts: What regulatory or infrastructure moves would actually move the needle for public services and national security without slowing us down?
Where This Leaves Us
The window for shaping this is narrowing. Every team that pastes sensitive data into a frontier model without thinking through the implications is making a choice, whether they realise it or not.
The National AI Plan is a start. The AI Safety Institute is operational. GovAI is being built. But the real work happens at the team level, in the decisions engineers make every day about which model to use, which data to send, and which controls to implement.
If your organisation is evaluating whether your Cloud and DevOps foundations are ready to support AI in production with proper sovereignty controls, our Agentic AI Readiness Audit evaluates exactly that, scoring your infrastructure, data pipelines, governance, and team capability across six dimensions.
For a broader look at how we help engineering teams build AI-ready foundations, see our full list of services.