AI and the End of Work

AI and the End of Work: Complement or Substitute?

The end of work is the policy question of our times. Every general-purpose technology in history has seen work not only survive technological change but thrive. But AI is unlike any new technology that has gone before it.

The question that matters — will work survive? — is one our institutions have yet to ask. If we don’t ask the question and prepare to manage the change already wrought, then the existing economic processes that distribute the economy’s income will cease to work. It is better that we continue to share the gains than fight over the losses, so we need to get ready.

We owe our current high living standards to technological change. And, and we would do well to remember how technological change hauled most of humanity out of poverty before we succumb entirely to AI anxiety.

Britain, the first industrial country, saw its economy grow 115-fold between 1700 and 2002[1]. Australia grew 4,000-fold. Canada — not ordinarily given to boasting — grew an astonishing 8,000-fold. Poverty-level subsistence for most of human history gave way, through successive waves of technological disruption, to the prosperity we now take entirely for granted.

But history also counsels patience. Electrification took seventy years to deliver its full economic promise. The ICT revolution took forty. Both created enormous dislocation along the way. The difference this time is speed. AI is being integrated into work processes before we have had time to absorb what it is, let alone what to do about it.

The Jevons Paradox

AI is a general-purpose technology, one that fundamentally changes every aspect of economic and social life, as the steam engine, railways, electrification, and information and communications technologies did before it. What distinguishes a GPT from a merely useful invention is its pervasiveness: it reaches into every corner of the economy and rewires how things get done.

English economist Stanley Jevons, born in 1835, showed us something that still holds: any technology that lowers the cost per unit of output, and becomes widely available, should produce more output at a lower price. Demand increases, and employment rises. This is the Jevons Paradox — the counterintuitive finding that efficiency gains tend to expand activity as demand rises rather than merely substitute labour for machine.

Lawyers are finding AI’s effects on hourly rates a challenge for law form partners as it disrupts the long standing sweatshop partnership model. Like educators, lawyers need to respond to an unstoppable technology and develop a new organizational model. We should anticipate lower-priced legal services will allow those currently excluded by price into the market for legal services. Even though prices fall, capturing lower priced volume could be a winner for those who grasp the Jevons Paradox.

Manufacturing unions were tamed in the 1980s are professional unions next? Don’t hold your breath when the regulators are primarily lawyers.

The fear of automation is nothing new. It has accompanied every major wave of technological change and has, so far, never been vindicated at the aggregate level. Technological change has created more jobs than it has destroyed, and new jobs that our predecessors could not have imagined. Total employment in existing job titles in the United States in 1940 was roughly 50 million. Many of those titles still exist and still employ around 50 million people — but another 100 million jobs now exist that no one in 1940, or even 1990, could have foreseen.

That is the record. But the record has its limits as a guide, and AI is testing them.

Two Forces, One Reckoning

The early inventions of the industrial revolution identified two effects that new technologies exert on the labour market. Economist Daniel Susskind calls them the complementing force and the substituting force, and the distinction is everything.

The Spinning Jenny, developed in 1760, transformed cottage-based yarn spinning into a factory industry, raising production, lowering prices, and stimulating demand. It was a complementing force — it made workers more productive, not redundant. The Jacquard loom of 1805 was something else entirely: a paper program fed into a loom that automated complex weaving patterns, allowing the unskilled to do the work of the skilled on a mass scale. The Jacquard loom was a substituting force.

The history of automation is a history of tension between these two forces and their effects on the labour market. Complementing tools — nail guns for carpenters, MRI machines for radiologists — augment the value of human expertise and raise the wages of those who use them.

Substituting tools — GPS for taxi drivers, to take a pointed example — commoditize expertise. The black cab drivers of London once spent years memorising the city’s maze of streets, sitting oral examinations to demonstrate they knew the most efficient route between any two points. That knowledge had market value. It no longer does.

We need to be honest about where we are. In the United States, wages in production occupations fell by almost 8% between 1960 and 2015 as manufacturing tasks were computerised and handed over to robots.

According to David Autor, general labourers saw wages decline by 6% per year as disruption took hold. Those subject to the complementing force — doctors, lawyers, financial professionals, technicians, executives, and managers — saw their wages rise between 2% and 6% annually. The labour market is not a neutral actor in the face of technological gains. It distributes the impact of change unevenly, and the net effect on both the quantity and quality of jobs, whether created or destroyed, is determined by which force dominates.

The AI reckoning is already underway. Recent research by Goldman Sachs shows that firms deploying AI as a substituting force — cutting costs, eliminating job postings — are doing exactly that. Firms pursuing the complementing force are capturing productivity gains and hiring. The Jevons Paradox is playing out in real time, but in two directions simultaneously. Old businesses need new people who can execute new tasks. Many companies are now hiring AI natives rather than investment bankers and CPAs, recognizing the future that the Jevons Paradox portends.

What Is Different This Time? The End of Polanyi’s Paradox.

Of the many tasks we execute, we do most of them effortlessly — and we cannot articulate how. We cannot define rules for how we tell a joke, what makes a good cook, and how a seasoned analyst smells a flawed argument before the numbers give it away. If we cannot specify the rules, we cannot program the task. Until now, Polanyi’s paradox has not been challenged.

However, AI is different because it learns autonomously from unstructured text. Large language models produce output that bears no resemblance to rules-based programming. AlphaFold determined the folding structure of proteins — a problem that had eluded structural biologists for decades — without being programmed with any rules for how to do it. Based on amino acid sequences previously beyond our ability to interpret, the implications for drug discovery are profound.

This phenomenon is genuinely new. AI can encroach on tasks that prior automation could not touch, across multiple occupations simultaneously. Economists call this task encroachment — the chipping away at processes within a given job, rather than the elimination of specific jobs altogether — and the breadth and depth of AI’s potential encroachment have no precedent.

The industry talks about “agentic AI,” a term designed to excite executives about substituting AI for headcount. But it is the wrong framing. A more accurate description is delegative AI: a tool that shortens the distance between intent and outcome, and one that will become as ubiquitous as Microsoft Office is today. With great respect for all who have helped me in the past, AI is the best research assistant in my career. An assistant that finds not just the source, but the paper, the paragraph, and the sentence containing that nugget of data and analysis to prove the hypothesis.

Whether it is truly intelligent is a separate question, and one worth contemplating. Give AI every piece of information available in the late eighteenth and early nineteenth centuries and it will not produce Einstein’s theory of relativity. It is, for now, a very sophisticated chauffeur: it will drive you to the data and the analysis, but it cannot create what is yet to be created.

The Policy Lag

Global polling by Ipsos finds that citizens in the United States, Canada, Britain, and Ireland are significantly more anxious about AI than their counterparts in Asia. Some observers suggest the anxiety is less about the technology itself than about what corporations will do with it — and with those whose jobs it displaces. That is the right question. It is also, conspicuously, the one that policymakers have yet to contemplate: the impact on the labour market at the macro level.

Labour markets are the mechanism through which national income is distributed. In aggregate, some 65% of national income flows to workers; 35% to capital. The labour market also determines what we contribute — through taxation — to the public services that underwrite our standard of living: health care, education, income support. If the substituting force dominates and labour income declines, those services may no longer be financed through conventional means. The fiscal arithmetic of a high-substitution AI transition is one that no government has yet attempted to model, let alone define a response to.

The answer is not to protect jobs, but to ensure the survival of work. Policy must guide AI so that work is preserved. We must have sustained investment in education and training, job transition income support, and a renewed commitment to competition policy that prevents the gains from AI from accruing exclusively to the hyperscalers and their shareholders.

Canada, which has observed what raw economic power can deliver when deployed without constraint, should understand better than most why that commitment matters.

And What of the Here and Now?

AI technology stocks are powering ahead. Unlike the internet crash of 1999 to 2000, equity prices are rising to capture earnings that have, so far, consistently outperformed expectations.

The AI infrastructure buildout is structurally similar to those faced by investors in late nineteenth-century railway expansion and the late twentieth-century internet infrastructure boom. The risk is that AI repeats the past and that revenues arrive too late to sustain the debt necessary to build it.

And there are markers of trouble, notably the self-funding dynamics between the hyperscalers and the AI hardware manufacturers. Some correction will come, leaving a capital overhang and debt that cannot be serviced. Much of the debt sits with investors rather than the banking system, which limits the risk of a financial crisis — but economic growth will take a hit as losses need to be distributed and taken.

The worst real economy outcome relative to expectations — observed by economic historian Adam Tooze — is that AI proves not to be a GPT but something more like the airline industry: a useful technology that builds a useful industry but on which investors find it difficult to make much money.

Expectations of the bounty from the universal adoption of AI are behind this year’s price momentum, but the muted market response to Nvidia’s recent upside earnings surprise suggests the bar to ever higher prices has been raised. Financial-market overshoots are common in momentum and carry trades, but they rarely end until expected gains give way to expected losses. The gap between them is likely narrowing.

Financial markets are discounting mechanisms, and they can offer a weighted wager on future economic outcomes. One signal that the market is placing more weight on the complementing effect than on the substituting effect as AI works its way through the economy would be a shift in the driving investment style — from momentum to value.

Stay tuned