Is AI creating common skills across jobs or deepening divisions?

The paper asks whether AI diffusion is producing "a common set of competencies across occupations or deepening occupational divisions," and answers that it does both at once. Using large-scale online vacancy data from ten countries across the Global North and Global South, it finds AI demand "overwhelmingly concentrated within a narrow technical core": roughly three quarters to four fifths of AI related vacancies sit in STEM occupations in every country. Its conclusion is that "AI is not homogenizing the labor market. It is splitting it."

Inside that core, AI intensive jobs "share a remarkable focus on Python, SQL, machine learning, and data analysis," which the authors describe as a "compact and highly transferable core of data competencies" that recurs across sectors, countries, and career stages. Outside it, convergence "remains limited" and mostly shows up as divergence of occupations. So the split is between technologically intensive occupations, which grow more alike, and the wider labor market, which grows less like them. The authors say this convergence is most strongly expressed at the point of labor market entry, so the shared data core appears to be an entry requirement more than a later-career specialization.

The method, as the abstract describes it, combines natural language processing and a large language model with multilevel bipartite network analysis, mapping links between occupations, required skills, and career stages. That three-way structure is what lets the paper say the pattern holds across countries and stages at once rather than in one aggregate.

This sits alongside existing work on how AI exposure is distributed, but at a different level. Does concentrated AI exposure enable workers to adapt and reallocate? finds concentration inside firms' task bundles that lets workers move to unaffected tasks. Here concentration appears across occupations and in advertised skills, and the paper reads it as structural bifurcation, not as a cushion. The two are not in conflict, since they measure different things (tasks versus vacancy skill requirements), but this paper offers no evidence on whether workers can move across the divide. The occupation-and-task map in What collaboration level do workers actually want with AI? starts from worker preferences; this one starts from employer demand.

The excerpt does not state the sample size, time window, country list, or how AI related vacancies were identified, and it gives no effect sizes. Vacancy postings record skills employers ask for, not skills workers use, hiring outcomes, or wages, so the excerpt does not show that the divide affects pay or employment. The entry-level emphasis also cannot say whether the shared core is learned on the job or brought in, which is where When does AI actually boost worker productivity? would bear if it held here. What the excerpt does support is narrower: as advertised, AI skill demand is neither a general upskilling wave nor a simple deepening of specialization, and analyses that average across occupations will blur a two-part structure.