The 2025 finding needs its original frame
This is a retrospective reading of the ILO and Poland’s NASK refined occupational-exposure index, published on 20 May 2025—not a new announcement. Its most quoted result was that one in four workers globally were in occupations with some exposure to generative AI. The same study made a less headline-friendly but more important qualification: because occupations contain tasks requiring human input, transformation rather than replacement was its most likely outcome. [2]
That distinction changes the question for employers and policymakers. “Exposure” does not mean an occupation disappears; it means that some of its constituent tasks may be susceptible to automation or material redesign. A role can therefore become more productive, more tightly monitored, narrower, better augmented, or worse in quality without being eliminated. The index was designed to locate where those possibilities warrant attention, not to declare a fixed future for workers.
The 2025 release arrived amid fast-changing model capabilities and rising familiarity with GenAI tools. The authors updated the ILO’s 2023 index accordingly. Its contribution was methodological as much as numerical: it shifted analysis toward the task composition of work and toward gradients of exposure, instead of presenting a simple automated/not-automated binary.
A task map rather than a verdict on occupations
The index drew on a representative sample from 29,753 tasks in the Polish occupational classification system. It collected 52,558 observations on perceived automation potential for 2,861 tasks from a survey of 1,640 employed people across one-digit ISCO-08 groups, then compared those inputs with expert surveys and Delphi-style discussions. The resulting knowledge base supported an AI assistant that predicted scores for task descriptions in ISCO-08 documentation. [1, 2]
This design matters because job titles conceal variation. Two people with the same occupational label can perform different mixes of documentation, customer contact, judgment, physical work, coordination and accountability. Conversely, a task that looks automatable in a description may remain difficult to deploy in practice because of workflow integration, data access, reliability requirements, regulation, cost, infrastructure or employer choice.
The index’s four progressively increasing exposure gradients were an attempt to retain that nuance. Clerical occupations remained the most exposed group, while some highly digitised professional and technical work in media, software and finance showed increased exposure. The latter point complicates the familiar story that only routine, lower-skilled work matters. Capability-based GenAI can affect cognitive, text-heavy and structured-information tasks as well.
Yet the index did not measure realised adoption, realised productivity gains, wages, redundancies or job quality. It measured potential exposure based on task descriptions and assessments. That boundary is not a technical footnote; it is the condition for using the figures responsibly.
The distributional signal was stronger than the headline
The global estimate combined broad reach with sharp differences. The study put 3.3% of global employment in the highest exposure category, with 4.7% for female employment and 2.4% for male employment. Overall exposure rose from 11% of employment in low-income countries to 34% in high-income countries. In high-income countries, the highest gradient covered 9.6% of female employment versus 3.5% of male employment. [2]
These are exposure shares, not projected dismissal rates. Still, they identify a distributional problem that generic “AI skills” programmes can miss. Where clerical and administrative work is concentrated among women, a task-level technology shock can be uneven even when aggregate employment remains stable. A firm that automates scheduling, formatting, data entry or drafting may redistribute oversight and exception handling upward, reduce entry pathways, or change workload and bargaining power. None of those outcomes follows automatically from the index; all are plausible implementation questions it helps surface.
The country-income pattern deserves the same restraint. Higher exposure in high-income economies may reflect their occupational structures and digital intensity. It does not establish that lower-income economies are insulated. Limited infrastructure and slower adoption can reduce near-term implementation while also constraining workers’ ability to benefit if deployment accelerates later. Exposure is an early signal about the work that could change under specified technological capabilities, not a ranking of national resilience.
What later evidence clarified
A later ILO brief, published on 17 April 2026, made the interpretive rule explicit: AI exposure indicators should not be read on their own as predictions of job losses or labour-market outcomes. It noted that such measures rely on static descriptions of current tasks, omit economic feasibility and adoption constraints, and contain subjective assumptions. Most fundamentally, they capture what AI could do as a first analytical step, not what will happen. [6]
That clarification does not invalidate the 2025 index. It defines its appropriate role. Exposure analysis is useful for triage: identifying occupations, task clusters and groups that deserve closer observation. It becomes insufficient when it is used to infer headcount plans, wage effects or social outcomes without complementary evidence.
A stronger evaluation chain would connect the index to actual workplace data: tool availability and use; task redesign; time allocation; quality and error rates; employment, hours and wages; internal mobility; training access; and worker experience. Comparing these measures before and after implementation—and against comparable work that was not redesigned—would help distinguish capability from consequence. The sources supplied here do not provide those causal results, so they should not be assumed.
Practical implication: govern the transition at task level
For organisations, the practical use of the index is not to produce a “jobs at risk” list. Start with a transparent inventory of tasks, including which ones can be assisted, which may be automated, and which require human review, accountability or relationship work. Then test changes in bounded workflows before redesigning roles or setting staffing expectations.
Workers and their representatives should be involved before deployment decisions harden. The 2025 study called for social dialogue and targeted policy responses; that is operationally relevant because workers often know where task descriptions diverge from real work. Their input can reveal hidden coordination, safety checks, client context and error-recovery work that a model-oriented assessment may miss.
Governments can use exposure maps to target training, transition support and labour-market monitoring—especially where gendered occupational concentration is high. But they should pair the maps with observed indicators of adoption and job quality. The right policy question is not “which jobs will AI erase?” It is: which tasks are changing, who controls that redesign, who gains time or income, and what safeguards ensure that transformation improves rather than degrades work?
Limitations
This archive analysis is constrained to the supplied ILO, NASK and republication captures. The principal 2025 evidence is institutional research, not an independent causal evaluation of workplace outcomes. The index used Polish task data and an ISCO-08-based approach to generate global estimates; its reported figures are exposure estimates, subject to methodological assumptions and differing national conditions. No claim here should be read as a forecast of future job losses, adoption rates or productivity effects.