A recently circulated piece from CitriniResearch imagined what an AI-driven economic crisis might look like, writing a fictional macro memo from June 2028 in which the S&P 500 has fallen 38% and unemployment has hit double digits. I have already published a detailed rebuttal explaining why that specific scenario fails on its own terms, from the incoherent “Ghost GDP” framing to the feedback loops that supposedly have no natural brake.
Dismissing the dramatic version, however, does not mean there is nothing worth worrying about. The Citrini scenario makes for gripping reading precisely because it touches a real nerve, even if the mechanism it describes is economically confused. What follows is an attempt to build a more realistic framework for thinking about how AI could create genuine economic friction, why that friction is a transition problem rather than a permanent one, and why the destination we are heading towards is unambiguously good.
The story of rising living standards is, at its core, a story about reducing the scarcity of important things. Every major leap in human prosperity has followed the same pattern: a technological breakthrough makes a critical resource relatively more abundant, its price falls, and what was once a luxury available only to the wealthy becomes accessible to everyone.
Agriculture is the clearest example. Before the mechanisation of farming, the majority of the human population spent the majority of its time producing food. A varied diet with reliable access to meat, dairy, and out-of-season produce was an aristocratic privilege. The series of agricultural revolutions from the 18th century onwards, culminating in the Green Revolution of the mid-20th century, reduced the share of the US workforce employed in agriculture from around 40% in 1900 to under 2% today. Food did not become free, but it became radically cheaper relative to incomes, and diets that were once the preserve of the wealthy became universal.
Industrialisation did the same for manufactured goods. Before the factory system, a well-made shirt or a piece of furniture was expensive enough that most households owned very few of either. Mass production collapsed the cost of physical goods to the point where a modern worker on a modest wage owns more material possessions than a medieval lord. In 1900, a bicycle cost the equivalent of several months’ wages for the average worker. By 1950, it cost a few days’.
In both cases, the pattern was the same. Technology made a scarce resource relatively more abundant. The relative price of that resource fell. What had been a luxury became ordinary. The economy adjusted. Living standards rose.
The engine that coordinates these shifts is the price mechanism. In an economy with many competing uses for scarce resources, prices act as signals, telling producers what to make and consumers what to buy, balancing supply and demand across thousands of interconnected markets simultaneously. When a resource becomes relatively more abundant, its price falls relative to other resources, and the economy rebalances accordingly. Capital, labour, and materials flow away from activities where returns have fallen and towards activities where returns are now relatively higher.
This is not a metaphor. It is the literal mechanism by which economies absorb technological change, and it has operated reliably across centuries of disruption. The important qualifier is “reliably,” not “painlessly.” The adjustment process works, but it does not work instantly, and the transition can be deeply uncomfortable for the people directly affected.
AI is the next breakthrough in this sequence. It reduces the relative scarcity of cognitive labour.
This deserves some unpacking, because “cognitive labour” is a broader category than most people instinctively assume. The natural reaction is to think of it as creative or analytical work, the kind of deep thinking that feels distinctly human. Some cognitive labour is indeed like that. Most of it is not.
Consider the actual task structure of a compliance analyst performing KYC checks at a bank: reviewing documents against a set of established criteria, flagging discrepancies, following a decision tree. Or a paralegal conducting document review: reading contracts, identifying relevant clauses, comparing language against templates. Or a junior accountant reconciling figures across systems according to well-defined rules. These are credentialed roles within prestigious institutions, often well-compensated, and the barrier to entry involves years of education and professional qualification. The work itself, however, is highly routine. It is pattern-matching against established rules, performed in natural language rather than on a factory floor, but structurally not so different from the kind of repetitive physical labour that machines replaced a century ago.
Entry-level software engineering is perhaps the most striking example. A junior developer writing CRUD endpoints, fixing standard bugs, and implementing well-understood patterns earns $100,000 to $150,000 in a major US city. The role is protected by a reputation for difficulty and by technical credentialing that filters heavily at the point of entry. Yet a large proportion of the day-to-day work is highly systematisable, and AI coding tools are already demonstrably capable of performing it. This is a role where the compensation reflects the historical scarcity of the skill far more than the complexity of the typical task.
None of this is to disparage the people doing these jobs. It is simply to observe that the white-collar economy contains far more routine cognitive work than the salary bands and job titles would suggest. When a technology arrives that can perform routine cognitive tasks cheaply and at scale, a very large portion of the labour market is affected.
If AI follows the same pattern as agriculture and industrialisation, the long-run implications are extraordinary. Just as mechanised farming democratised access to nutrition and mass production democratised access to material goods, AI has the potential to democratise access to cognitive services.
Today, a good lawyer costs $300 to $1,000 per hour. Bespoke software tailored to a small business’s specific needs costs tens or hundreds of thousands of dollars. A personal financial adviser is a service largely reserved for the affluent. Personalised tutoring, one of the most effective educational interventions known, is accessible only to families who can afford $50 to $150 per hour for it.
AI changes the economics of all of this. Not by making human lawyers or tutors worthless, but by making the cognitive labour component of these services relatively more abundant and therefore relatively cheaper. More people gain access to legal guidance, to custom-built tools, to financial planning, to education tailored to their specific needs. This is the same mechanism that made food and clothing affordable, applied to a new category of scarcity. It is, on any reasonable accounting, very good news.
It is also worth being precise about what AI does and does not change. AI does not make human labour less capable than it was before. A lawyer does not become worse at lawyering because an AI can also draft a contract. Labour remains scarce. There is no shortage of useful things for people to do. What changes is that there is now a new, efficient partial substitute for some of what cognitive labour does, which means the relative price of that particular kind of work should fall. Total productive capacity goes up. The economic pie gets bigger. In principle, no one needs to be worse off.
“In principle,” however, is doing a lot of work in that sentence.
Here is where things get uncomfortable, and where a realistic concern about AI’s economic impact differs from the dramatic crisis scenario.
The real price of routine cognitive labour should fall to reflect its new relative abundance. In a textbook, this happens smoothly: wages adjust, workers reallocate to tasks where humans retain a comparative advantage, and the economy reaches a new equilibrium. In reality, nominal wages are sticky downward. This is one of the most robust empirical findings in labour economics, documented across countries and decades. People resist pay cuts. Employers know that cutting nominal wages destroys morale and triggers the best employees to leave first. Contracts, minimum wages, and social norms all reinforce the floor.
When the market-clearing real wage for a particular type of work has fallen but the nominal wage cannot adjust downward to reflect this, the result is a quantity adjustment rather than a price adjustment. Employers do not cut pay; they cut headcount. The economic term for this is involuntary unemployment, and it is the standard Keynesian account of how labour markets malfunction when prices cannot move freely.
This is not speculative. It is the mechanism that has operated in every deflationary or disinflationary episode in modern economic history. The question is not whether this dynamic exists, but whether the AI-driven shift in relative prices will be sharp enough and fast enough to trigger it at a scale that matters.
If it does, the likely policy response is well-established and, frankly, not very exciting. Central banks and governments facing rising unemployment reach for expansionary tools: lower interest rates, fiscal stimulus, tolerance of higher inflation. The purpose of tolerating inflation in this context is not mysterious. If the real price of cognitive labour needs to fall by, say, 20%, there are two ways to get there. One is for nominal wages to fall 20%, which workers and institutions fiercely resist. The other is for nominal wages to stay flat while the general price level rises 20%, which achieves the same real adjustment without anyone’s payslip showing a smaller number. Inflation acts as the lubricant for a relative price adjustment that the economy cannot otherwise make cleanly.
This is essentially the argument Tobin made in 1972 and Akerlof, Dickens, and Perry formalised in 1996: that moderate inflation “greases the wheels” of the labour market by allowing real wage adjustments that would be impossible in a world of perfectly stable prices. It is also, not coincidentally, the playbook that central banks actually deployed after 2008, when they held interest rates near zero and tolerated above-target inflation for years to ease the adjustment from the financial crisis.
There are several reasons to think the negative scenario, while internally coherent, may not materialise in practice.
The first is that AI capability and AI adoption are very different things. The capability frontier is advancing remarkably fast, and there is good reason to expect it will continue to do so. The integration of that capability into actual business processes, however, involves procurement cycles, regulatory approval, organisational change management, IT integration, retraining, and the sheer institutional inertia of large organisations that have been doing things a certain way for decades. History suggests that transformative technologies take far longer to diffuse through the economy than their enthusiasts expect. Electricity was commercially available in the 1880s; it did not meaningfully transform factory productivity until the 1920s, because the gains required redesigning the entire layout of manufacturing around the new power source rather than simply swapping electric motors into existing configurations.
If AI adoption follows a similar pattern, the displacement of cognitive labour may be gradual enough that the economy can adjust organically through attrition, differential hiring, and natural career transitions, without the sharp unemployment shock that triggers the inflationary policy response.
The second reason is that new industries and roles tend to emerge alongside the displacement, though they are always invisible in advance. It is a cliché to note that nobody in 1990 predicted that “social media manager” or “app developer” would be significant job categories, but clichés endure because they are true. The economic logic here is straightforward: when the cost of cognitive labour falls, activities that were previously too expensive to undertake become viable, and entirely new categories of economic activity become possible.
The third reason is that comparative advantage still holds. Even as AI handles an expanding range of cognitive tasks, scarcity shifts to its complements: judgment, taste, trust, physical presence, the ability to take responsibility for a decision, the capacity to navigate genuinely novel situations where no established pattern exists to match against. This is not wishful thinking about the irreducible specialness of humans; it is the mathematical consequence of opportunity cost, as Ricardo described it in 1817. If AI’s relative advantage is largest in routine cognitive work, rational allocation means deploying it there, which necessarily leaves domains where humans retain a comparative edge.
The important caveat is that reallocation, while it works in aggregate, is historically very painful for the specific people caught in the transition. The comparison to manufacturing decline is instructive. The US economy created plenty of new jobs in the decades after 1980, but the workers who lost manufacturing jobs in specific regions often did not get them. Whether the AI transition is manageable depends heavily on its speed, its breadth, and the degree to which it affects career entry versus mid-career displacement. A 50-year-old compliance officer cannot easily retrain as a robotics technician, regardless of how many retraining programmes the government funds.
The economics of AI are, on balance, clearly positive. A technology that makes cognitive services cheaper and more widely available follows the same pattern as every previous breakthrough that raised living standards. The price mechanism will eventually coordinate the adjustment. Total output rises. The pie gets bigger.
The real uncertainty is whether we manage the transition competently. The economics will sort themselves out, given time and reasonable policy. Whether the politics provide that time is the genuinely open question. If governments are too slow to deploy the tools they have, if populist anger leads to counterproductive intervention, if the adjustment is fast enough to generate visible pain but not fast enough to generate visible benefits, the transition will be more costly than it needs to be.
This is a more modest claim than the AI doomers make, and a more honest one. The risk is not that the economy breaks. The risk is that we fumble a transition we have every tool to manage well, and that people suffer unnecessarily in the interim. That risk is worth taking seriously precisely because it is realistic rather than cinematic.