Explainer: What is sovereign AI and why does Australia lag in this area?

Discussions in Australia are heating up around the need for “sovereign artificial intelligence (AI).”

According to a report by the CSIRO, the country lags behind others in AI research and development. And this, experts warn, is putting us at the back of the pack in using AI to boost productivity and bolster the economy.

Australia’s AI lag

The CSIRO notes that foundation models – the kind of machine-learning (ML) algorithms trained on vast amounts of data which powers products such as OpenAI’s ChatGPT – are predominantly made overseas.

At least 125 foundation models have been developed worldwide in recent years. Of these, most are made by private tech companies. 73% come from the US, 15% from China and most of the rest from Europe.

Anton van den Hengel is a professor of computer science at the University of Adelaide and is the founding director of the university’s Australian Institute for Machine Learning (AIML).

He tells Cosmos that the vast majority of “what needs doing in any economy using AI is efficiency gains.” Implementing AI in industry will make goods and services cheaper, making them more competitive, he adds. This is a process which van den Hengel says could take another 20 years.

The rest of AI’s impact is new applications which don’t yet exist and can only be realised through research. “A nation having deep sovereign AI capability is critical to it being able to capture all of those economic opportunities,” van den Hengel says.

“We had quite a reasonable, competitive AI research sector globally,” he adds. Emphasis on “had.”

“We’ve got some very good software engineers, but they’re not really ML engineers. And we’ve got a very small number of very good researchers, but we miss almost everything in between.”

What’s the point of sovereign AI?

“We’ve all been impressed by the way these models can write a wedding speech or a poem. But the speed, power, and colossal scale of the data analysis they can achieve has the potential to help us solve our greatest challenges, boost productivity and save lives,” says lead author of the CSIRO’s report, Dr Stefan Hajkowicz.

“A foundation model for healthcare for example could help us untangle complex, hidden relationships in patients’ health records, helping us reduce the 140,000+ [human-error] medical misdiagnoses in Australia each year.”

Van den Hengel adds: “The easiest kind of infrastructure that you could point to would be a large language model for Australia.”

“Everybody says, well why don’t we just download the American model? It’s not a million miles away.”

According to van den Hengel, it’s not that simple and such an approach could cause problems.

“These large language models form the core infrastructure that all other AI is built upon. If you use somebody else’s model, then everything else carries their value judgements, their intonations and their priorities.

“What does it mean for our national identity if our school kids are being educated using a system that thinks that Joe Biden is the head of our democracy?”

CSIRO scientists agree.

“While there are significant benefits to fine-tuning existing models in terms of cost and the speed of innovation, using foreign models poses security and reliability risks,” says Professor Elanor Huntington, the CSIRO’s Digital, National Facilities & Collections Executive Director. “It may also result in tools that aren’t culturally appropriate in an Australian context, or that don’t realise the benefits for our workers that we want to see.”

A matter of education

Van den Hengel argues that research is critical to having AI-enabled education and industries in Australia.

“If you want more AI-enabled teachers, or you want more AI high school teachers, or you want more AI lectures in universities, you can’t just go and you can’t print them. You’ve got to train them. That means that you need trainers. You can’t just print the trainers either. Somebody’s got to train the trainers.

“Those graduating PhD students do research and then they train some PhD students of their own, who train PhD students of their own. That pool of PhD students then teaches in the university and train the undergrads. The undergrads then go and train the teachers and the teachers train the high school students.

“But the information has to come from somewhere. And where it comes from is PhD graduates. And the way you get more PhD graduates is to fund research. We are in the unenviable position of starting at least 10 years late in this process.”

Van den Hengel says that all of Australia’s output of PhDs in AI is equivalent to that of a single medium-sized university in the US.

Where to for Australian AI?

“Australia has done very well, in some focused areas of research. We’re very good in computer vision, and have been for quite a while. And computer vision’s interesting because that’s the bit of AI that really drove the revolution,” van den Hengel explains.

He also notes that Australian researchers have strong skills in natural language processing and robotics.

“There’s about 5 or 6 really world class groups in Australia.” But, van den Hengel says they’re “still not really operating at the scale of groups overseas.”

He believes that Australia has a chance to turn things around.

“This is our chance to improve the productivity of every industry, and to really build a globally competitive AI based economy for kids are unfortunately at the moment we’re going backwards.”

Sign up to our weekly newsletter

Please login to favourite this article.