Artificial intelligence has diffused through the economy faster than any previous technology, yet the labour market carries almost no trace of it. Unemployment across the advanced economies remains near record lows. The predicted white-collar contraction has not materialised yet. Trying to solve this contradiction between doomsday predictions and unaffected markets can help identify AI’s economic pain points.
The first clue is in how the technology is used. Three detailed studies — from OpenAI, Anthropic and Google — converge on the same finding. OpenAI‘s analysis of 1.5 million conversations shows users overwhelmingly asking for information and help drafting, rather than delegating whole tasks. Anthropic’s data on Claude points the same way. Google‘s ATLAS study, the broadest to date, is also the most precise. AI appears in more than two-thirds of occupations, but in only about a fifth of the tasks within them. Fewer than one interaction in ten completes a task from start to finish. Thus, the technology, at present, is wide but shallow. It augments far more than it automates, which is why aggregate employment has moved so little.
That resolves parts of the puzzle, but also relocates the rest of it. If AI is not, for now, disrupting jobs, the policy-relevant questions concern not the level of employment but two dynamic properties of the adjustment. The first is how quickly labour markets absorb a productivity shock. The second is how its gains are divided. Each can unfold in more than one way.
The First Question: The Path of Adjustment
Consider what a productivity-raising tool does to the demand for a skill. It sets two forces against each other. AI can substitute for a worker, performing tasks the worker used to do. It can also complement the worker, raising output per hour (marginal productivity of labour) and the associated market value. Which force dominates depends on two conditions.
The first is how completely AI can perform the task unaided. The more self-sufficient the tool, the stronger the substitution. The second is whether cheaper, faster output expands the market or meets a ceiling. Where demand for the output is elastic — where lower costs bring in many new customers — the complementary force tends to win, and employment and wages in the field rise. Where demand is capped, the same productivity gain lets fewer workers meet unchanged demand, and employment falls. This is the same logic underlying previous patterns. Mechanisation raised agricultural yields enormously, yet shrank farm employment, because the world’s appetite for food did not expand in tandem.
The first scenario branch is already visible in practice. Wherever AI primarily assists and output can grow, demand for workers strengthens. Where AI largely replaces and output cannot grow, demand weakens. Most occupations sit between these poles, which is why the aggregate numbers look placid although the underlying composition is shifting.
However, the favourable branch carries a timing problem. Suppose demand for a skill rises and wages climb. Workers respond, but slowly, because acquiring a skill takes years, while the wage signal is immediate. The economist Richard Freeman showed in the 1970s how this lag produces cycles. A demand surge for engineers sent wages up, and students poured in. The cohort that graduated years later met a saturated market and falling pay, which deterred the next cohort. The market did not converge smoothly. It overshot, repeatedly. Law degrees and coding bootcamps have since traced the same arc.
Figure 1: The mechanics of a productivity jump in a competitive market
AI sharpens this dynamic. The same tools that raise productivity also lower the cost of acquiring the skill, through AI tutoring, coding and writing assistants. The delayed wave of entrants can therefore be larger, and arrive faster. The heavier the inflow, the deeper the subsequent glut. The incidence is also uneven. The adjustment falls first on new entrants, whose footholds are the codifiable, junior tasks AI performs most readily. A Stanford study finds employment for workers aged 22 to 25 in the most exposed occupations down about 16 percent relative to older colleagues. The evidence is contested. The Yale Budget Lab sees no broad disruption yet, and others date the weakness to before ChatGPT. The theoretical mechanism, though, predicts this concentration of risk at the foot of the ladder.
The Second Question: The Division of Gains
Higher productivity does not, by itself, reveal who ends up better off. That depends on bargaining power along the chain, and, decisively, on competition among AI suppliers. Here too there are two scenarios.
In the first, AI remains a competitive, commoditised input. This describes the present. The price of a given level of AI capability has fallen at extraordinary speed, by some estimates tenfold a year, as rival labs and open models undercut one another. When AI supply is competitive, its gains disperse widely, to firms, workers, and consumers alike.
In the second scenario, the frontier consolidates. Building leading models is costly, and concentrated among a few firms. A stable oligopoly could price by a different rule: charging by the value of the human labour (value of marginal product) it displaces instead of the computation costs (marginal cost). This is called value-based pricing. Signs of it already exist as premium tiers are priced far above their compute cost. Under this scenario, the outcome reverses. Workers grow more productive, yet the surplus is captured upstream as the proprietors’ profit, rather than passing to workers or their employers. A case of rising output and stagnant wages, with the gains accruing to a handful of suppliers, is not a labour-market problem at all. It is a modern version of the age-old rent distribution conflict.
Figure 2: Scenarios under different assumptions about demand elasticity and AI market structure
Policy implications
Earlier technologies eventually generated new kinds of work, and AI is expected to do the same, so the market eventually finds a new equilibrium. The societal costs, however, arise in the transition — unevenly, and without correcting themselves. Several countries have previously implemented policies, built for earlier waves of automation and trade, that provide replicable frameworks. Those discussed below map onto one of the mechanisms above, but they need to be fine-tuned for AI.
On the skilling front, there are multiple precedents. Singapore’s SkillsFuture gives every adult stackable, modular credits and even mid-career skilling, allowing requalification to be completed in months. Germany’s Qualification Opportunities Act funds retraining for workers who are still employed, even during short-time work, rather than targeting only those who are unemployed. Sweden’s collectively bargained job-security councils step in early on after dismissal and replace workers quickly, even having provisions for entrepreneurship. To integrate the AI challenge, these programmes need to be coupled with real-time labour-market information from vacancy data so that vocational training can track live demand.
Protecting the entry rung has precedents too. Austria’s apprenticeship guarantee, and the wider EU Youth Guarantee, promise every young person a job, apprenticeship or training place within months. For displacement, America’s Trade Adjustment Assistance offered wage insurance that topped up the pay of re-employed workers and lifted their earnings by roughly a quarter. This design is worth reviving but only after restructuring. It covered workers over 50 displaced by trade, whereas AI falls first on the young, so it should now target early-career cohorts and reward firms that keep training-intensive junior roles rather than automating them.
On the rent-seeking front, the tools are relatively new. The EU’s Digital Markets Act can impose interoperability, data access, and easy switching on designated cloud gatekeepers. Britain’s competition authority has already extracted commitments from the largest cloud providers to cut exit fees and ease switching. Public compute, such as the US National AI Research Resource, the EU’s AI Factories and India’s own IndiaAI programme, offer smaller players an alternative to renting from a handful of firms. Improvements are still needed to extend switching and interoperability rules up from the cloud to the models layer, to make scrutiny of cloud-and-model tie-ups routine rather than case-by-case, and to incorporate multi-vendor requirements into public procurement.
Thus, none of the labour-market apprehensions stem from the technology alone. The speed of the adjustment and the distribution of its gains are ultimately policy variables. The economies that act fast and push policy in the right direction can turn this disruptive transition into a managed outcome.
Arya Roy Bardhan is a Junior Fellow with the Centre for New Economic Diplomacy at the Observer Research Foundation.
This article was originally published on the Observer Research Foundation website.