Greens plant their flag on AI policy, but the portfolio is unrealistic
The Greens are right that Australia's AI governance is broken — but their solution reveals a misunderstanding of what it actually takes to build a frontier AI model.
Senator David Shoebridge's appointment as the Australian Greens' spokesperson on Digital Rights and AI is the kind of move that sounds more consequential than it is. The Greens are right that AI governance in Australia is fragmented, right that foreign-controlled infrastructure creates real dependencies, and right that the technology deserves a dedicated policy framework. Where they come unstuck is in the part of the pitch that sounds boldest: the idea that Australia can build "ethical Australian owned and controlled AI" to meaningfully push back against the frontier labs.
Australia's regulatory gap is real, and the Greens are right to name it
The regulatory critique is solid. AI governance in Australia is genuinely scattered across the Online Safety Act, the Privacy Act, sector-specific regulators and a queue of consultations that have been running long enough to become a genre. The Greens' call for a dedicated AI Act and a Minister for AI is not an eccentric position. The European Union has its AI Act. The United Kingdom has its AI Safety Institute. Canada has proposed its Artificial Intelligence and Data Act. Australia is behind on institutional architecture, and Shoebridge is not wrong to say so.
The sovereignty argument is also worth taking seriously on its own terms. As The Bearing has noted in its coverage of the News Bargaining Incentive negotiations, the structural relationship between Australian institutions and US platform companies consistently favours the platforms. When every layer of the AI stack runs on foreign-controlled infrastructure, that dependency is real. Data sovereignty, regulatory jurisdiction, and the question of who can actually enforce decisions against a company headquartered in San Francisco are legitimate concerns, not confected ones.
But the policy that flows from that diagnosis is where the logic breaks.
The frontier model race is an arithmetic problem, not a political one
Training a frontier AI model is not a question of national will or clever engineering. It is a question of capital at a scale that simply does not map onto Australian public budgets or private markets. OpenAI's GPT-4 training run cost somewhere between fifty and one hundred million US dollars, in early 2023 - a lifetime ago in AI terms. And that figure is now the floor, not the ceiling, for frontier model development. Google, with revenues of nearly three hundred and fifty billion US dollars a year and one of the world's largest privately held compute infrastructure, is not clearly winning the frontier model race against OpenAI, SpaceXAI and Anthropic. China, with unlimited state resources and a coordinated national program, has produced genuinely impressive models but has not displaced the American frontier labs either.
The idea that Australia, with a research sector and public investment capacity that neither approaches those scales, will build a model competitive with the leading frontier labs does not survive contact with the arithmetic.
The idea that Australia, with a research sector and public investment capacity that neither approaches those scales, will build a model competitive with the leading frontier labs does not survive contact with the arithmetic.
This is not a critique of Australian research talent, which is genuinely strong. Australia has produced world-class AI researchers, many of whom now work at the labs the Greens are proposing to compete with. The problem is not people. It is compute, capital, and the compounding advantage that comes from being first to train on the largest datasets at the largest scale.
What Australia can actually build is more modest and more useful
What Australia can plausibly build is something more modest and more useful: domain-specific models trained on Australian data, tuned for Australian legal, medical, agricultural, and administrative contexts, running on infrastructure that is jurisdictionally legible. That is a defensible and achievable goal. It would deliver genuine public value. It is a much quieter ambition than "push back against US dominance," but it is the one that the evidence supports.
The Greens are also proposing to examine whether the government's Duty of Care legislation adequately addresses AI chatbots and agents. That is a reasonable legislative test to apply. As The Bearing's coverage of the Fix Our Feeds algorithm bill showed, the mechanics of how platforms shape behaviour matter more than the framing around them, and the same principle applies to AI agents operating in Australian digital environments.
The frontier model race is being run between organisations with effectively unlimited compute budgets and years of compounding advantage. Australia is not going to win it. A government that was honest about that constraint, and focused its resources accordingly, would do considerably more for Australian AI capability than one chasing a flag that is already too far down the track.
Sources
The Bearing — Tech giants hold upper hand in media negotiations
The Bearing — Algorithm opt-in is really a platform opt-out
Frequently Asked Questions
Why can't Australia build its own AI to compete with ChatGPT or Gemini?
Training a frontier AI model costs somewhere between tens and hundreds of millions of dollars per run, and those costs are rising with each new generation. Australia's public research budgets and private capital markets are not operating at the scale required — and even Google, with revenues approaching three hundred billion US dollars a year, is not clearly ahead of OpenAI in the frontier race.
What is the Australian Greens' AI policy?
The Greens have appointed Senator David Shoebridge as their spokesperson on Digital Rights and AI, and are calling for a dedicated AI Act, a Minister for AI, and the development of 'ethical Australian owned and controlled AI.' The platform also proposes examining whether existing Duty of Care legislation covers AI chatbots and agents.
Does Australia have an AI regulatory framework?
No dedicated framework exists. AI governance in Australia is currently spread across the Online Safety Act, the Privacy Act, and a range of sector-specific regulators, with a series of consultations still underway. By contrast, the EU has an AI Act, the UK has an AI Safety Institute, and Canada has proposed its own dedicated legislation.
What kind of AI could Australia realistically build?
The realistic opportunity is in domain-specific models — AI trained on Australian data and tuned for specific contexts like law, medicine, agriculture, and public administration. This is a narrower ambition than competing with the US frontier labs, but it is achievable and would deliver genuine public value.
Why does it matter if Australia's AI infrastructure is foreign-controlled?
When every layer of the AI stack runs on foreign-controlled infrastructure, Australia's ability to enforce its own regulations against companies headquartered overseas becomes structurally limited. Data sovereignty, regulatory jurisdiction, and the practical enforceability of decisions are all affected — these are the same dynamics that have consistently favoured US platform companies over Australian institutions in earlier tech policy negotiations.