Australia's AI Strategy Needs Lateral Flexibility

Why Australia's AI Strategy Needs Lateral Flexibility

A case for modular re-education and a fairer distribution of AI's gains. (Assisted by Claude)

The gap in an otherwise good speech

On 15 July, Anthony Albanese stood at the University of Sydney and gave a substantial speech: 'AI in Australia's interests'. It involves a new Office of AI inside the Department of the Prime Minister and Cabinet, a national standards framework, real protection for artists and journalists against having their work strip-mined for training data, rules for data centre energy and water use. Measured against where the debate was even a year ago, it's a serious document, and the general reception has been positive.

But read it closely and there's a hole in it. Education gets one sentence: the Education Minister "meeting with his counterparts... to discuss the impact of AI in schools." That's it. Schools. Not the electrician who's about to find their TAFE-taught trade partially automated, not the paralegal whose document review job has quietly evaporated, not the radiographer watching an AI system read scans faster and, increasingly, better than they can. When the University of Sydney gathered its own experts to respond to the speech, not one of the eight raised adult reskilling or labour transition as a subject in its own right. One, Dr Mike Seymour, noted pointedly that the speech said nothing about "increased funding for R&D, basic research, or universities." The silence is conspicuous.

This matters because the Prime Minister's own framing gives away the stakes. He told the story of starting as a bank teller talking customers into trusting an ATM, and noted that "that world is long gone... Australians studying here or starting in the workforce today don't expect – or want – one job for life." He's right — the age of the butcher, the baker, and the candlestick maker, one trade learned young and held for life, is long gone too, and it isn't coming back. But an economy that no longer offers one job for life needs an education and training system that no longer assumes one either. Ours still does. And that mismatch, not some speculative rogue-AI scenario, is the risk actually sitting in front of us.

The certain disruption versus the hypothetical one

I've just finished Sebastian Mallaby's The Infinity Machine, his biography of Demis Hassabis and DeepMind. It's a useful corrective to a debate that spends a lot of its energy on the wrong kind of fear. The existential threat from AI — the Terminator scenario, the paperclip maximiser, the sudden treacherous turn — remains exactly what it has always been: a hypothesis, seriously argued by serious people, but a hypothesis nonetheless, resting on chains of assumption about future capability and motivation that nobody can currently verify. Reasonable people disagree about how much weight it deserves, and I'm not going to pretend to settle that here.

What isn't hypothetical is the effect on how people earn a living. That part is already happening, and it doesn't require artificial general intelligence to keep happening — it just requires AI to keep getting incrementally better at the narrow, structured, pattern-based components of jobs that make up most of the labour market: reading images, drafting documents, writing code, answering routine queries, coordinating logistics. The Department of Employment and Workplace Relations analysis the PM cited in his own speech confirms the shift is already showing up in the data, even while headline employment numbers stay healthy. Good news in aggregate can still mean real disruption in the particular — for the individual radiographer or paralegal, an aggregate unemployment rate near historic lows is cold comfort if the specific skill they spent a decade building has just been devalued.

So the argument I want to make sits deliberately apart from the debate about rogue AI. It doesn't need existential risk to be true. It only needs employment disruption to be real, which is no longer seriously contested by anyone — including, evidently, the government's own labour market analysts.

What "lateral flexibility" means, and why the system doesn't have it

Australia's education and training pathways are built like a series of separate, purpose-built buildings rather than a shared set of components. If you want to be a doctor, you sit the UCAT, get into a six-year medical degree (or a graduate-entry program after an undergraduate degree), and complete years of supervised training. If you're already a nurse — with years of clinical experience, physiology, pharmacology, patient care — and you decide you want to become a doctor, the system mostly doesn't ask what you already know. It asks you to start again, near the beginning, alongside eighteen-year-olds.

That's not a hypothetical example. It's the standard experience of career-changers in regulated professions across Australia, and it generalises well beyond medicine: teachers moving into instructional design, tradies moving into renewable energy installation, retail managers moving into logistics technology. In each case, the destination job overlaps substantially with the skills the person already has. In each case, the system is structured to make them prove they don't have those skills by making them sit through the material again.

My proposal — and I don't think it's original so much as it's an obvious idea whose time is overdue — is to think of qualifications as sets of Lego bricks rather than sealed boxes. If becoming a doctor requires modules 1 to 100, and becoming a nurse requires modules 60 to 120, then a nurse who wants to become a doctor needs modules 1 to 59: the bricks they don't already have, not the ones they already do. It's the difference between buying a whole new Lego set to build a crane when you already own the castle set, and simply buying the crane add-on pack that shares half its bricks with what's already on your shelf.

There's one design detail that matters more than any other: the destination credential itself doesn't change. Our nurse-turned-doctor doesn't graduate with some lesser, hybrid, fast-tracked qualification — they sit the same final licensing exam, on the same day, marked to the same standard, as every other medical graduate. What's compressed is the road to that exam room, built from credit for what they already demonstrably know; but the exam itself, and the degree it confers, stay exactly as they are. That's what makes the idea credible rather than a shortcut, and it's the detail that does the most work later in this essay, when I get to why employers don't trust alternative credentials.

The mechanism that would make this real is fourfold: competency-based assessment on the front end (you're tested on what you know, not on time served in a classroom), portable, machine-readable records of exactly which modules you hold, genuine recognition of prior learning that isn't a bureaucratic afterthought but the front door of the system, and — critically — an unmodified, shared terminal assessment at the end, so the qualification you finish with is identical to the one everyone else finishes with.

Has this been tried? Yes — and the results are instructive rather than conclusive

This isn't a blank-page idea. Australia has already, on paper, endorsed something close to it.

The 2024 Australian Universities Accord — the most significant review of the tertiary system in decades — explicitly called for "more modular, stackable and transferable qualifications," recommended that microcredentials be funded and formally accredited within the Commonwealth-supported place system, and proposed a National Skills Passport: a single, portable record that would let someone demonstrate exactly what they know to any employer or institution, backed by "regularised recognition of prior learning." That is, almost verbatim, the Lego brick idea. It's sitting in a government report right now, mostly unimplemented, while the AI strategy conversation happens in a different room.

Internationally, the most mature real-world version is Singapore's SkillsFuture Credit, running since 2015. Every citizen 25 and over gets a training credit (topped up further for mid-career workers) that they can spend on modular, bite-sized courses. It's genuinely stackable — course credits build toward larger qualifications rather than sitting in isolation — and it's evaluated seriously: a national survey (TRAQOM) tracks outcomes at both the course level and six months later, alongside sectoral labour market indicators. In 2019, 86% of over 43,000 surveyed trainees said they could perform their job better after training. That's a real, running example of a government treating modular reskilling as core infrastructure rather than an afterthought — and it's worth Australia studying in detail rather than reinventing.

The module system only works if two supporting structures are built alongside it and taken as seriously as the modules themselves: a credentialing standard employers actually trust (which the shared-exam principle largely resolves, and which the National Skills Passport is meant to formalise across the system if it's ever properly funded and built), and genuine, enforceable recognition of prior learning, not the current situation where Recognition of Prior Learning (RPL) exists in policy but is routinely too slow, too discretionary, or too costly in practice to be worth pursuing. Singapore's investment in outcome tracking is instructive here too — you can't build employer trust in a credential system without also building the evidence that the credentials mean something.

A second disruption: can we still trust the exam itself?

There's a complication underneath all of this that deserves its own airing. A teacher at a private school recently described to me how her Year 9 and 10 boys had become expert at using AI not just to draft essays and assignments, but to prompt the tools to disguise the fact — "make my essay look as if it weren't produced by AI." That's not a classroom management problem; it's a preview of a second, quieter disruption sitting underneath the one this essay is mostly about. If AI can produce coursework indistinguishable from a student's own work, then take-home essays, portfolios, and continuous assessment — the bulk of how most qualifications, including many microcredentials, are actually assessed — stop being reliable evidence of what a person knows.

This cuts two ways for the argument here. It's a genuine threat to the credibility of any credentialing system, modular or traditional, that leans on unsupervised coursework. But it also strengthens the specific design choice above: supervised, sit-down, high-stakes exams — the kind medical boards, bar exams and trade licensing tests already use — are comparatively hard for a language model to sit on a candidate's behalf. Keeping the terminal assessment unmodified and invigilated isn't just about preserving employer trust in a stacked pathway; in an AI-saturated world, it may be one of the few assessment formats left that still reliably measures the person rather than the tool they had open in another tab. If Australia is rethinking assessment for the AI era at all — and given what that teacher described, it clearly needs to be — the pathway architecture and the integrity-of-assessment problem should be designed together, not treated as separate conversations.

The other half of the problem: who gets the gains

Modular reskilling addresses one half of the disruption — helping people move sideways into new work. It does nothing about the other half: what happens to the share of national income going to labour if AI and, increasingly, humanoid robots start competing directly with human wages across a widening range of jobs, while a smaller group of capital owners captures a growing share of the value created.

Sam Altman's 2021 essay "Moore's Law for Everything," is a good one to build on because of how deliberately it's framed. Altman's argument isn't the standard progressive case for redistribution — it's explicitly pitched as pro-growth and pro-business, on the theory that a wealth-transfer argument alone won't survive contact with a tax-paying electorate that (rightly) worries about disincentivising the investment that creates the wealth in the first place. His mechanism: tax the two asset classes that will capture most of AI's value — corporate equity and land — at a modest, fixed annual rate (he proposes 2.5% of each, phased in as GDP grows), and distribute the proceeds directly to citizens as cash and equity, not as a means-tested welfare payment. The land tax piece draws straight from Henry George: land value rises mostly because of what society builds around it, not because of anything the owner did, so it's fair for society to recapture some of that value. The equity piece is designed so that everyone — not just shareholders — has a direct stake in company share prices rising, which Altman argues changes the politics: people who own a slice of the gains are less likely to vote to strangle the thing producing them.

That framing is worth borrowing regardless of what you think of the specific 2.5% figure. The reason UBI proposals keep running into resistance in Australia isn't that people reject the idea of a safety net — it's that a payment funded from general income tax reads, to a lot of working taxpayers, as their wages being redirected to people who aren't working, which is a much harder sell than a payment funded from a new tax on the assets AI itself is inflating. If AI genuinely does what its advocates claim and collapses the cost of goods and services the way computing costs collapsed under Moore's Law, then a modest, transparent claim on the resulting corporate and land value — phased in, indexed to actual productivity gains rather than promised in advance — is a more defensible ask than raising income tax to fund a payment to people whose jobs a robot just took.

Australia already has two pieces of the institutional plumbing this kind of scheme needs, and neither had to be invented from scratch. Compulsory superannuation is, in effect, a mechanism the country already trusts for building broadly-held capital ownership over a working life. And the Future Fund shows Australia is comfortable with the state holding productive assets on the public's behalf and paying out the returns over time. A future-facing "AI dividend" doesn't need to be sold as a foreign concept — it can be sold as an extension of instruments Australians already understand and, on the whole, like.

Putting it together

The Prime Minister is right that Australia has form on getting ahead of disruptive change rather than just absorbing it — Medicare, universal super, the social media ban. The AI speech shows the government is willing to apply that instinct to sovereignty, copyright and data centres. It hasn't yet applied it to the two things that will determine whether ordinary Australians experience this transition as an opportunity or a loss: whether they can move sideways into new work without starting from zero, and whether they get a fair share of the wealth AI generates even if their own job is one of the ones it takes.

Three concrete tasks follow from this, and all three build on work the government has already half-started rather than requiring it to invent something new:

First, fully fund and build the National Skills Passport the Universities Accord already recommended, with teeth: mandatory employer and institutional recognition of stacked modules, a genuine RPL fast-track rather than a discretionary one, and outcome tracking modelled on Singapore's TRAQOM survey so the system can prove — not just promise — that its pathways mean something. Put the "modules 1–100, shared terminal exam" logic explicitly into the design of at least a few major regulated-profession pathways (health, teaching, engineering) as a pilot, with patient- and public-safety review built in rather than assumed away.

Second, treat assessment integrity as part of the same reform, not a separate one. As AI makes unsupervised coursework an increasingly unreliable signal of what someone actually knows, invest in supervised, high-stakes assessment infrastructure — for schools as much as for professional pathways — as the trust anchor the whole modular system depends on.

Third, open a serious, public conversation about a capital-and-land-funded dividend, phased and modest to start, explicitly framed as a claim on AI-driven productivity growth rather than a welfare payment — building on the institutional trust Australians already have in compulsory super and the Future Fund, not asking them to accept an entirely foreign idea.

None of these requires resolving the rogue-AI debate. All three are worth doing whether the existential risk turns out to be nothing or something. That, I think, is exactly the case worth putting to government: not "AI might end humanity," which is contestable and distracting, but "AI is already changing who can do what job, whether we can still tell what anyone actually knows, and who captures the resulting wealth" — none of which is contestable, and none of which the current strategy has much to say about.