PARIS—The debate over artificial intelligence and employment has become preoccupied by a false choice. One side predicts large-scale job destruction; the other notes that aggregate data still show little. Both observations can be true. While AI’s economy-wide employment effects remain limited, the conditions for disruption are already emerging at the level of tasks, firms, and early careers. Given this, waiting for clear evidence that the shock has arrived would be a mistake. The data that eventually settle the argument may arrive only after the opportunity to shape the outcome has passed.
The productivity promise is no longer merely theoretical. In a study of 5,172 customer-support agents, access to a conversational assistant increased the number of problems resolved per hour by 15%. The least experienced workers improved the most, suggesting that the tool was diffusing some of the tacit knowledge of top performers. But beyond the model’s capability frontier, AI can worsen performance by producing answers that are coherent, plausible, and wrong.
Task-level gains do not translate automatically into economic growth. They shrink when machine output must be checked, coordinated, and integrated into a product or service. Over the next few years, the main constraint on AI-driven productivity growth is likely to be organizational absorption rather than model capability. Like the advent of electricity, AI requires firms to redesign processes and responsibilities before its full benefits appear.
Caution is equally important for employment. As of August 2026, payroll data covering millions of US workers reveal no evidence of widespread economy-wide displacement. Yet employment among 22-to-25-year-olds in AI-exposed occupations is now 19% below where it would be if it had kept pace with employment among their less-exposed peers. The gap reflects reduced hiring more than increased dismissals.
The finding reveals a flaw in the way the issue is usually framed. As Luis Garicano of the London School of Economics has shown, a job is not a task; it is a bundle of tasks. For experienced professionals, technical expertise, client knowledge, judgment, accountability, and coordination tend to form a tightly connected whole. AI complements that work. A junior role is often a looser bundle. Research, initial analysis, first drafts, or simple coding can be detached and automated. The same technology can therefore raise the value of the experienced professional while hollowing out the junior job.
The most serious near-term risk is not mass unemployment, but the breakdown of apprenticeship. Professions have produced experts in much the same way for centuries: beginners perform elementary tasks and learn by doing. Those are precisely the tasks AI absorbs first. In a small randomized experiment conducted by Anthropic, developers using an AI assistant scored 17 percentage points lower on a test of mastery than those working without one, while the time saving was not statistically significant. AI is most useful to people who already possess the expertise needed to check it, even as it threatens the route through which that expertise is acquired.
This contradiction suggests that public policy must focus on rebuilding the first rung of the career ladder. Apprenticeships and academic programs should be judged by the skills students acquire, not by whether graduates can complete tasks that machines can now perform. Employers should give young people ownership of complete projects and test their ability to explain, challenge, and correct automated output. Because this training imposes a private cost while producing a large social return, it will require incentives and public co-financing.
The risk begins even before jobs disappear. AI use remains concentrated among highly paid and highly educated workers, turning a technology with equalizing potential into a multiplier of existing advantage. This pattern is not inevitable: employer-sponsored training is among the strongest predictors of workplace adoption. The durable skill is not mastery of a particular interface, but the ability to frame a problem, verify sources, detect errors, and know when human judgment must prevail.
Governments must also measure change before it appears in aggregate unemployment data. Relevant indicators include the share of vacancies open to beginners, changes in required credentials, AI adoption by income level, and labor’s share of value added in AI-intensive sectors. Each indicator should be tied in advance to a threshold for action. A dashboard without trigger points would merely document delay.
Firms, meanwhile, face a choice between two forms of deployment. They can automate fragments of an existing process, accumulating pilots, reviews, and disappointments. Or they can redesign the process from end to end, specifying what may be delegated, what must be checked, and what requires human judgment. This choice determines both productivity and the quality of work. The direction of the technology is not fixed: competition policy, public procurement, liability rules, and worker participation can favor the augmentation of human capabilities over simple substitution.
Europe should not try to avoid disruption at the cost of falling further behind. Slow adoption would sacrifice productivity without preserving jobs for long. Europe must accelerate adoption while giving workers the time and institutions needed to adapt.
On September 17, Philippe Aghion and I are bringing economists, technologists, business leaders, and policymakers together at the Collège de France to clarify these choices. The aim is not to manufacture consensus, but to separate disagreements that better evidence can resolve from decisions that require a collective judgment.
AI can broaden access to expertise, support innovation, and restore the growth that Europe lacks. But there is no guarantee that its benefits will spread on their own. Shared prosperity will depend less on the performance of the next frontier model than on our ability to preserve skill formation, reorganize work, and distribute the gains equitably. We still have time to do so. The absence of an aggregate shock is not a reason to wait; it is the reason we can still act.
Copyright: Project Syndicate, 2026.
www.project-syndicate.org
What if Africa invested its assets and wealth differently?