Sovereign AI is only as good as the data behind it

Article appeared in the Herald Sun on 9 September 2026.

You wouldn’t use a GPS if you knew the map might be wrong. Yet businesses are using AI to guide decisions without always knowing whether the underlying data is accurate or where the model learned its version of the world.

The Hon Dr Andrew Charlton MP, Assistant Minister for Science, Technology and the Digital Economy, addressed the broader question of AI sovereignty in a recent speech at the ANU Crawford School of Public Policy, Why Australia needs to turn compute into sovereign AI.

Charlton argued that Australia has a significant opportunity in AI, but that simply hosting data centres and supplying the energy to run them is not enough. If Australia remains at the infrastructure end of the value chain while overseas companies own the chips, models, software and intellectual property, he warned, we risk becoming a permanent “importer of intelligence” while much of the economic value created by AI flows offshore.

He also raised an important question about Australian data. From agriculture and resources to the creative industries, Australia possesses valuable knowledge and information that will help train and improve AI systems. We need to think carefully about who gets access to that data, on what terms, and who ultimately benefits from the value it creates.

But there is another dimension to AI sovereignty that deserves much more attention: whether we can actually trust the data those systems rely on.

Many organisations know their data is imperfect. The challenge is stopping those flaws from being carried into reporting, analytics and AI systems. 

That starts with how data is captured and governed at the source.

Systems need to be designed with clear definitions, validation rules and accountability for data quality. Too often that work is skipped because the focus is on getting a new system implemented on time rather than thinking about how the information it captures will ultimately be used.

The challenge becomes harder again when organisations are relying on data from third parties. They do not control the quality or provenance of that information, yet it may still feed increasingly consequential decisions. 

That is why proper data governance is essential. Organisations need common rules and definitions applied consistently, as well as a clear understanding of how poor-quality data can affect the decisions that follow. 

AI layered over poor-quality data doesn’t just reproduce existing problems. It can amplify them.

And with large language models (LLMs), there is another layer of uncertainty.
Organisations may have limited visibility into the data used to train a model, how current that information is, which sources were included and what biases may have been embedded along the way. 

Small language models (SLMs) can give organisations greater control. They can be built on existing open source models and then fine tuned or refined using an organisation’s own data for a much more specific purpose. 

But that only brings the question of data quality closer to home. If an organisation’s own information is inaccurate, inconsistent or poorly governed, a smaller and more controlled model can still produce unreliable outputs.

Before leaders allow AI to inform consequential decisions, they should be asking: what data sits behind this? How reliable is it? Do we understand the limitations of the model? And, most importantly, have we validated the output before acting on it?

Robodebt offers a useful warning.

It was not an AI system in the way we use the term today, but it showed the danger of treating systematically generated outputs as inherently reliable. Automation does not remove the need for judgement or accountability.

Australia should absolutely be thinking about sovereign AI and how we retain more of the economic and strategic value created by our data.

But sovereignty without trust is not enough. Control over AI systems only creates value if we can trust the data they rely on and the outputs they produce.