Australia: The regulation superpower

There are risks to over-regulating ai. by david hughes.

First published in the MRC’s Watercooler newsletter. Sign up to our mailing list to receive Watercooler directly in your inbox.

Australia is a world leader in ‘regulation’, as our Prime Minister likes to remind other world leaders, whether at the UN or a climate conference. What a strange thing to gloat about. It suggests his proudest achievement is the list of restrictions Australia places on its people and businesses. Are we really so reckless that we need the world’s toughest regulations to thrive? 

Governments have a long habit of regulating things before they understand them. In 1865, the British Parliament passed a law requiring every vehicle to be led down the road by a man carrying a red flag. This law was created for steam engines, by politicians who had never even seen a car. When the car arrived 20 years later, the law was applied to the new technology and stayed until the turn of the century. Meanwhile, the Germans used those years to build motor industries. The British gave the world the railway and remained stuck in that age. 

I think there's a fair chance we're about to repeat the mistake on artificial intelligence. I'd be surprised if many MPs could explain how the technology works, and that's no criticism, because most people can’t. But that's a big deal because we have a parliament set to make laws for a machine it hasn't built, has barely used, and would rather slow down than study. 

I understand why many readers find the whole subject of AI unsettling. The internet only carries what we put into it. AI can write, reason, and argue back. People worry it will upend the way we live, and that's a real concern. Much of how we live is worth preserving, and any sensible approach to regulating AI should start from that premise. But hesitation on its own protects nothing. 

This technology will be at the centre of our children’s and grandchildren’s lives, whether Australia is ready for it or not, and nothing we decide this year can stop it. Being ready means we can build it here, we have a say in how it is built, and we keep the best of the way we live now.

So what are our politicians actually deciding? On the table is around $150 billion of data centre investment by 2030. Before the money comes, investors want answers to three questions, and the government has answered none of them.

Where the power comes from

AI doesn't live on your phone or laptop. Every question you ask travels to a data centre, a warehouse full of computers, and the answer comes back. A large data centre draws as much electricity as a small city. It needs that power every hour of the day, in all weather. Engineers call this firm power. Wind and solar only work when the weather allows, so batteries, gas, or coal have to fill the gaps. The energy market operator expects data centres to use about 6% of the grid by 2030, and around 12% by 2050. 

The government's predictable answer is a renewable energy mandate. Every large data centre must bring enough new renewable generation to match 100% of its demand. In the US, the same companies are signing contracts to bring on new nuclear generation and rapidly expanding their gas-powered generation capacity to power data centres. Here they'll be told to build wind farms. A sensible solution would make them pay for the power they use and the poles and wires that carry it, and let them choose where it comes from.

Copyright: who gets paid

An AI model learns by reading. To build one, a company feeds it an enormous quantity of text, pictures, and recordings, and the model works out the patterns in all of it. Most of that material is somebody's property: books, news articles, songs, photographs. The question is whether a company can read all that without asking, or has to get permission and pay.

I heard a media executive say this week that “we're about to hand over our national identity to foreign companies”. That’s far too dramatic. The media companies would sell what they call the national identity tomorrow for the right price. The people with a real claim are the writers, musicians and journalists whose work these models are trained on, and they deserve credit and payment. The government ruled out a free-use exception a year ago and promised a licensing scheme instead. 

A year on, there is no such scheme. One of the largest developers said it wants to train models here but can't license content from thousands of small rights holders one at a time. It’s not as complex a problem as it sounds. Radio stations have paid a single collective license for the music they play for decades, so the solution isn't novel. But we still await a decision. 

Whether AI is trained here

AI models are built through training. It is the reading I described above but done at a massive scale: a warehouse of specialised chips running for months. Using the finished model, which is what you do when you ask a chatbot a question, is called inference. Almost every data centre in Australia does only that. Training is where the knowledge and the leverage are. A handful of American companies and their Chinese rivals are building the most capable models, and it’s vital that the democratic Western world stays ahead. 

This is not just an economic question. The same frontier AI models that draft an email can also analyse intelligence and test the security of systems critical to our sovereignty. A country that can train its own models strengthens its sovereign capability. If we rely on AI models built and owned elsewhere, we hand that capability to someone else and must trust they will keep sharing it. The US and UK now treat advanced AI as central to their defence, and we must keep pace with our allies.

The government accepts all this in its consultation paper, then proposes that the companies building those models be "granted authorisation" to train here, on conditions yet to be written. Investors are already looking elsewhere while the complex rules are being written.

Being a world leader in regulation is a strange ambition. We will likely see our Prime Minister stand up at a summit and tell the room that Australia now has the world’s toughest rules on artificial intelligence. Albanese fulfills his lifelong desire and becomes the modern version of the man walking in front of the car with the red flag.

What he won’t say is that the $150 billion earmarked for Australia went to Texas and Malaysia instead. Or that the best Australian engineers followed the work. Or that we lost the security of having access to the best models. At the end of the day, no amount of regulation will stop Australia's avid use of AI. We’ll just be using somebody else’s, built under somebody else’s rules, and we'll pay a premium for it.