The AI Abyss

Every previous wave of technological disruption has, in hindsight, resolved itself into a fairly clear story. The mechanisation of agriculture displaced farm labour and, over a generation or two, created industrial employment. Mechanisation of manufacturing pushed much of the workforce into the services economy, creating jobs offering employees greater variety and agency. The computer displaced clerks and typists and, over a generation or two, created an information economy that employs vastly more people, at vastly higher wages, than the jobs it destroyed. In every case, the destination was uncertain at the time and obvious in retrospect. What made these transitions bearable, even in their most painful moments, is that few seriously doubted humanity would still have economically valuable work to do once the dust settled. The only real question was what kind of work, and how long the adjustment would take.

Artificial intelligence may or may not follow that pattern. That “may or may not” is the entire point of this essay, and I want to be honest from the outset that after several months spent building a detailed economic model of how an AI-driven productivity shock would move through a modern economy, I am not in a position to resolve it. What I can do is explain, in plain language, why the uncertainty is so severe, what the range of plausible outcomes looks like, and why I think the most useful thing anyone can currently say about the economics of AI is not a prediction but a description of what has to go right.

I have called this essay “The AI Abyss,” and the title needs some justification, because it isn’t meant to be gratuitously alarming. When people reach for historical comparisons to help them think about the scale of AI’s disruptive potential, the one that recurs most often is nuclear weapons — another technology that arrived faster than institutions could adapt to it, carrying genuinely civilisational stakes. But the comparison is imperfect in an important way. Nuclear weapons presented an essentially binary question: they would either be used at civilisational scale, or they would not. Everything of consequence hinged on that single branch. AI is different. Something consequential is coming — that part is close to certain, given the pace of capability improvement over the past several years and the scale of capital now being committed to it. What remains almost entirely open is the shape of what’s coming: whether it is broadly benign or broadly destructive, whether it unfolds over years or decades, and whether the institutions responsible for managing the transition are remotely equal to the task. An abyss, in the sense I mean it, is not a prediction of catastrophe. It is an acknowledgement that we are looking into something whose depth and character we genuinely cannot see, and that treating it as either a nothing-event or a foregone triumph is a failure of intellectual honesty in both directions.

This matters for the broader argument of this project for two connected reasons. First, AI is, for a great many people, a significant and legitimate source of psychological discomfort — not free-floating technological anxiety, but a reasonable response to genuine uncertainty about their economic future and their children’s. Second, and this is the theme I want to draw out most strongly, the degree to which that discomfort is warranted depends heavily on whether governments can respond with a speed and creativity that, on the evidence of recent decades, they have not shown themselves capable of. I have written elsewhere in this project about the ossification of governance systems — the sense in which political institutions have become resistant to the kind of adaptive redesign that markets impose on firms through creative destruction. AI is likely to be the sharpest test that ossification has yet faced.

Three Camps, All Overconfident

Public commentary on AI’s economic consequences has, in my observation, settled into three broad camps, and I think all three are guilty of a similar sin: mistaking a plausible scenario for the scenario.

The first camp is dystopian. It pictures mass technological unemployment, a rapid and unmanaged collapse of the labour market’s capacity to absorb displaced workers, and possibly worse — up to and including loss-of-control scenarios in which AI systems themselves become the threat rather than merely the disruptor. The second camp is the orthodox economic reaction: AI is just the next in a long line of general-purpose technologies, no more world-defining than the steam engine, the electric motor, or the personal computer, and history’s pattern of eventual absorption will reassert itself as it always has. The third camp is techno-optimist: AI as a benevolent, near-universal solvent for scarcity, ushering in an age of abundance in which most work becomes optional and material wellbeing becomes close to unlimited.

Each of these camps has something right and something dangerously incomplete. The dystopians are right that the scale and speed of the change could genuinely overwhelm the mechanisms that have historically cushioned technological transitions — but they tend to underweight the extent to which policy choices, not just technological trajectory, will determine the outcome. The orthodox economists are right that historical technology transitions have, in aggregate, worked out far better than their contemporaries feared — but they tend to wave away the very real arguments for why this transition could be different in scale and speed from anything in the historical record. The techno-optimists are right that AI-driven abundance is a coherent and physically plausible scenario — but they are, in my view, dangerously light on the demand-side mechanics of how that abundance actually reaches ordinary households, and prone to gesturing at solutions like “universal high income” without engaging with how such a thing would actually be paid for, administered, or politically sustained.

I built an economic model specifically to try to get past the shouting match between these camps and put some actual structure around the argument. It didn’t resolve the disagreement. But it clarified, I think usefully, exactly where the disagreement lives.

Modelling an Economy Like a Web of Purchases

Before getting into what the model found, it’s worth explaining — without equations — what kind of model it is, because the mechanics matter for understanding why the results are structured the way they are.

The simplest way to think about a modern economy is as an enormous, interlocking web of purchases. Every business is both a seller and a buyer: a bakery buys flour, electricity and labour, and sells bread; the flour mill buys wheat and sells flour; the wheat farmer buys diesel and fertiliser and sells wheat. My spending is someone else’s income, and their spending is someone else’s income again, rippling outward through the economy in a chain that economists call the multiplier. This interconnectedness means that a shock to any one part of the economy — a productivity improvement in trucking, say — doesn’t stay contained to trucking. It changes the cost of everything that trucking touches, which changes the cost of everything those things touch, and so on. Government statistical agencies actually measure these purchasing relationships in enormous detail, in what are called input-output tables — essentially a giant matrix recording how many dollars of inputs from every industry go into producing a dollar of output in every other industry. I built my model on Australia’s official input-output tables, covering 115 separate industries, as a reasonably representative stand-in for a typical wealthy economy.

The model asks a specific question: if AI and robotics allow every industry to produce the same output using significantly less labour and capital — a productivity improvement, in economists’ language — what happens to the rest of the economy as that change ripples through the web of purchases? I modelled three scenarios of increasing intensity, corresponding loosely to modest AI augmentation of existing work, substantial but uneven AI deployment over the medium term, and something close to comprehensive substitution of cognitive labour across the economy. In the most intensive scenario, this amounts to more than doubling the economy’s underlying productive efficiency.

A productivity improvement of that scale sounds unambiguously good, and in a narrow technical sense it is — more can be produced with less. The entire difficulty, the reason this deserves an essay rather than a one-line reassurance, is that “more can be produced with less” does not automatically translate into “everyone is better off.” Whether it does depends on what happens to the people whose labour and capital are no longer needed to produce the old level of output, and on what happens to the income that used to go to them.

The Surprisingly Benign Headline — And Its Fine Print

When I first ran the model in its fully specified form, with what I’d consider reasonable central assumptions, the headline result surprised me by how closely it tracked the optimistic, orthodox economic story. Real economic output rose substantially in every scenario I modelled — by roughly a fifth in the mildest case, and potentially doubling the size of the economy in the most intensive case. Employment, remarkably, held up: it fell only modestly in the moderate scenarios, and actually rose in the most AI-intensive scenario I modelled. That is a genuinely striking finding, and if it is the correct one, it means the labour-market fears driving much of the public’s AI anxiety are substantially overstated.

But — and this is the crux of the whole exercise — that benign outcome is not something the model simply produces on its own from AI capability. It depends critically on a specific mechanism functioning as economic theory says it should, but cannot guarantee that it will: the emergence of substantial new industries, employing people in occupations that barely exist today, to absorb the labour that AI releases from existing work.

In the moderate scenario, roughly one in eight workers in my modelled economy would need to be employed in genuinely new kinds of work for the benign outcome to hold. In the most intensive scenario, it’s more than half the workforce. To be clear about what that means: the existing industries of the economy, even after accounting for all the additional demand that lower prices and higher incomes generate, simply do not employ enough people to replace what AI substitution removes. The slack has to be taken up by industries that, in a meaningful sense, do not yet exist.

Is this plausible? Historically, yes, emphatically. Streaming entertainment, app development, social media management, medical aesthetics, large-scale mental health services — these are genuine industries employing millions of people worldwide, most of which were negligible or non-existent thirty years ago. Rising incomes and falling costs have reliably, over the long run, generated demand for things nobody previously knew they wanted, and entrepreneurs have reliably found ways to supply them. This is precisely the mechanism that resolved every previous wave of automation anxiety, from the Luddites onward.

But “historically, yes” is not the same as “guaranteed, this time.” Three considerations give me pause, and I think they deserve to be taken seriously rather than waved away by either side of the debate.

The first is that having a comparative advantage is not the same as earning a decent living from it. Basic trade theory guarantees that human labour retains some economic role even when AI is better at literally everything, in exactly the same way that a country with no advantages at all can still benefit from trade with its more productive neighbours. But the theory says nothing about the wage that role commands. If AI can perform almost every task at near-zero cost, the price of the tasks only humans can do — genuine presence, embodied skill, authentic connection — might still fall to levels most people would not consider a viable living, even while nominally retaining “comparative advantage.” Horses, after mechanisation, retained a comparative advantage in a handful of niche applications for decades. That did not stop the horse population from collapsing, because the applications that remained were not enough to sustain anywhere near the previous population at a comparable standard of living. It’s an uncomfortable analogy to draw about human labour, and I draw it advisedly, but the economic logic is exactly parallel and needs to be reckoned with rather than dismissed as alarmist.

The second is that demand for new human-provided activities might itself saturate. People do not have infinite appetite for services, however novel; there is a point at which additional income increasingly flows into savings and asset accumulation rather than new consumption, particularly once the basics — health, education, food, shelter, entertainment — are comprehensively met. My model actually builds this saturation dynamic in explicitly, and it is one of the main reasons that in the most extreme scenario, sufficient headroom for new industries still exists — but the further into hypothetical abundance one pushes the scenario, the shakier the assumption that unlimited new demand simply keeps appearing becomes.

The third, and to my mind most pressing, is speed. The transitions that history points to as broadly successful — agriculture to industry, industry to services — played out over multiple generations. That gave labour markets, education systems, and social institutions decades to reallocate people gradually, largely through natural workforce turnover rather than mass involuntary displacement. If AI capability continues to improve at anything resembling its recent pace, the equivalent transition could be compressed into years rather than generations. None of our existing institutional machinery for retraining, relocating, or otherwise reallocating a workforce was designed to operate at that speed.

Who Gets the Gains

Even setting aside the employment question entirely, there is a second finding from the modelling that I think deserves more attention than it currently gets in public debate: regardless of what happens to the number of jobs, the share of total income flowing to labour, as opposed to the owners of capital, falls substantially in every single scenario I modelled, under every combination of assumptions I tested.

This is not a forecast of workers becoming poorer in absolute terms — real wages actually rise in most scenarios, sometimes substantially. It is a statement about relative shares: capital income grows two to three times faster than labour income across the board, because AI substitutes far more readily for labour than for capital, and the capital deployed to build and run AI systems itself commands a growing claim on the economic pie. In the most intensive scenario I modelled, the share of national income flowing to wages and salaries falls from roughly half to under a third — a shift with no real precedent in the economic history of any advanced, developed economy.

This matters enormously for how the whole story plays out, because of a mechanism that is genuinely counterintuitive but, once explained, quite intuitive. Most people, when their income rises modestly, spend most of the increase. Wealthy people and large capital holders, when their income rises very substantially, tend to save a much larger fraction of it — they have already met their consumption needs many times over. This is not a moral judgement, just an empirical regularity about human behaviour at very high income levels, and it turns out to be one of the single most consequential parameters in the entire model.

If the owners of the vastly increased capital income generated by an AI-driven economy spend it — on goods, services, experiences, whatever — that spending flows back into the economy as demand, sustaining exactly the kind of consumer-facing new industries needed to keep the whole system employed. If, instead, they save a large share of it — parking it in financial assets, offshore investments, or accumulating wealth rather than recirculating it — that money largely exits the productive, employment-generating part of the economy. In my modelling, this single behavioural assumption is the difference between an economy where employment rises modestly and an economy where employment collapses by more than half, even though the underlying technological capability and productive potential are identical in both cases. The economy’s capacity to produce grows in both scenarios. Its capacity to actually deliver that production to people who need it, through jobs and income, depends almost entirely on what the beneficiaries of AI-driven productivity choose to do with their gains.

Recent economic history gives some grounds for concern here rather than reassurance. The share of income flowing to labour, relatively stable across most of the twentieth century, has already been eroding gradually since around 1980, alongside the first waves of computerisation and globalisation. Large-scale monetary interventions like quantitative easing after the global financial crisis provided a natural experiment in what happens when large sums of money are injected into the hands of asset holders: financial asset prices rose dramatically, while the flow-through to broad-based real economic demand was comparatively modest. Neither of these is proof of what will happen with AI, but neither offers much comfort either.

Landing Somewhere in the Middle

I said at the outset that I would not pretend to resolve the uncertainty this essay describes, and I want to hold to that. What the modelling has genuinely done is clarify the structure of the question, not answer it. But clarifying the structure of a question is not nothing, and I think it is honest to say that the exercise pushes me toward a position that sits between the three camps described earlier, rather than validating any one of them outright.

It seems to me very unlikely that AI turns out to be a “nothing-burger” in the sense the orthodox camp sometimes implies — a productivity story no more consequential than previous general-purpose technologies, playing out on a similarly gentle multi-generational timescale. The scale of the productivity shift under even moderate assumptions, and the speed at which underlying AI capability has been improving, both argue against a business-as-usual outcome. Equally, I don’t think a rapid arrival of genuine post-scarcity abundance is a credible near-term prospect — the demand-side mechanics required to actually distribute the gains of an AI-driven economy broadly are far more fragile and far more dependent on active human and institutional choices than the abundance narrative generally acknowledges, and if those choices are not made deliberately, the gains concentrate rather than diffuse. At the other extreme, a full economic dystopia — mass unemployment with no offsetting adjustment at all — also seems to me the less likely outcome of the range on offer, precisely because the historical pattern of new industry creation and demand response has enough going for it that betting against it entirely also seems unwise. I want to be careful here though to separate the economic dystopia scenario from the question of existential risk from AI systems themselves acting outside human control — that is a genuinely distinct question, on which this economic modelling has nothing useful to say either way, and I do not think anyone honestly can currently rule it out.

Setting the existential question aside, then, my honest reading of where the balance of evidence points is toward an outcome that is highly consequential — a transition of a magnitude genuinely without precedent in the modern economic era — and one that is plausibly benign in its eventual, settled state, but only if it is well managed. And “well managed” is doing a great deal of work in that sentence, because even in the most favourable scenarios my modelling produces, the transition period itself looks genuinely difficult: substantial and rapid shifts in who works where, a marked change in the distribution of income between labour and capital that will strain the legitimacy of existing tax and welfare arrangements, and a narrow, contingent set of behavioural and policy conditions that need to hold for the benign outcome to be realised rather than the harsher alternative sitting immediately adjacent to it in the model’s parameter space.

Why This Belongs in a Project About Governance

This is, on the surface, an essay about economics, but it belongs in this project for a reason that has nothing to do with input-output tables. The central argument I have made throughout this project is that government’s core purpose should be the sustainable wellbeing of its citizens, and that the world’s governance institutions are poorly equipped to deliver on that purpose because they have calcified against the kind of adaptive, creative response that changing circumstances demand.

AI is likely to be the single sharpest test that calcification faces in the coming years, precisely because of what the modelling shows: the difference between a benign and a harmful outcome is not primarily a question of the underlying technology. It is a question of whether capital income gets recycled into demand rather than hoarded, whether new industries are allowed and encouraged to emerge fast enough to absorb displaced workers, whether tax and welfare systems adapt to a genuinely new distribution of factor income, and whether all of this happens on a timescale measured in years rather than the multi-decade timescales that have historically allowed institutions to catch up with technological change at a comfortable pace. Every one of those conditions is a matter of active government policy choice, not a matter of technological destiny.

Judged against the recent record of governance described elsewhere in this project — an inability to reform entitlement systems that everyone agrees are unsustainable, a demonstrated preference among incumbent political elites for protecting existing arrangements over adapting them, decision-making timescales measured in years for problems that may not afford years — there is a real and uncomfortable gap between what the AI transition plausibly requires of government and what government has recently shown itself capable of delivering. This is, I think, the honest root of a great deal of the public anxiety around AI, and it is a well-founded anxiety. It is not really a fear of the technology. It is a fear, grounded in reasonable observation, that the institutions responsible for managing its consequences are not up to the task.

I don’t think that fear should tip into fatalism. The point of building this model was never to generate a prediction; it was to clarify what needs to be true for things to go well, so that the case for building institutions capable of delivering those conditions can be made with some precision rather than vague reassurance. That case — for governance capable of speed, coherence, and genuine adaptive creativity — is the subject of the rest of this project.