Global AI Exposure Index Reveals 3.3 Percent of Workforce Faces High Automation Risk

In May 2025, the International Labour Organization published Working Paper 140, revealing that exactly 3.3 percent of global employment falls into the highest category of exposure to generative artificial intelligence. The 71-page document updates the agency’s 2023 estimates by incorporating multimodal model capabilities.[1]

The research team, led by Pawel Gmyrek and Janine Berg, partnered with Poland’s National Research Institute to evaluate 29,753 distinct occupational tasks. They collected 52,558 data points from 1,640 employed individuals to ground theoretical automation risks in daily workplace realities.[1]

Rather than treating occupations as monolithic entities, the researchers analyzed jobs as a "bundle of tasks," isolating the specific physical and cognitive actions workers perform. This granular approach prevents the overestimation of automation in roles that require complex physical dexterity alongside cognitive processing.[1]

The resulting index categorizes occupations into four progressive gradients of exposure, moving away from binary predictions of job replacement. The data indicates that transformation, rather than outright substitution, represents the most likely impact for the vast majority of the global workforce.[1]

The Polish Task Database

To build a more precise model than the standard international classifications allow, the researchers utilized Poland’s six-digit occupational system. This framework expands the standard 423 international categories into 2,541 specific occupations, detailing nearly 30,000 individual tasks.[1]

Poland serves as a representative baseline for the upper threshold of automation potential. The nation ranks between high-income advanced economies and emerging markets, possessing widespread internet access and digital infrastructure that make theoretical automation practically feasible.[1]

The Ministry of Family, Labour and Social Policy in Poland maintains this classification database, which Statistics Poland uses for labor market analysis. The research team scraped the public registry to extract job descriptions and map them directly to the International Standard Classification of Occupations.[1]

Professionals account for 29.3 percent of the tasks in this expanded database, while technicians and associate professionals represent 20.0 percent. Clerical support workers, despite facing the highest automation risks, account for only 2.6 percent of the distinct tasks cataloged.[1]

Human Baseline Calibration

The researchers initially deployed three large language models—GPT-4, GPT-4o, and Gemini Flash 1.5—to generate synthetic automation scores for every task. The models assigned a value from zero to one, where one indicated a task could be completed entirely without human participation.[1]

Because algorithmic predictions often diverge, the team designed a human-based evaluation to calibrate the models. They selected a representative sample of 2,861 tasks and distributed them to 1,640 Polish workers, ensuring that tasks were assessed by individuals employed in the relevant occupational groups.[1]

The survey sample was balanced by gender, with 51 percent female and 49 percent male respondents. Participants evaluated the susceptibility of 35 random tasks within their field on a scale from zero to 100, generating the foundational dataset for the revised index.[1]

While 71.2 percent of respondents reported having heard of generative artificial intelligence, actual workplace adoption remained low. Exactly 50.1 percent of the surveyed workers stated they never use the technology, and only 4.8 percent reported regular daily or weekly use.[1]

Expectations of technological disruption varied sharply by profession. Clerical support workers and professionals reported the highest expectations of moderate to significant impact, while manual and technical groups largely anticipated no impact or only slight changes to their daily routines.[1]

Expert Validation And Arbitration

To refine the human baseline, the researchers convened a panel of international specialists from the ILO and the Polish Ministry of Family, Labour and Social Policy. This expert group reviewed a semantically clustered sub-sample of 608 tasks to correct for worker biases.[1]

The experts consistently lowered the automation scores for physical tasks compared to the survey respondents. They highlighted the practical limitations of automating actions that require manual dexterity, such as operating injection molding machines or performing mechanical processing of silicon.[1]

Conversely, the panel raised the scores for structured, data-driven tasks within clerical and professional roles. The experts recognized that generative models excel at processing information and generating text, making tasks like drafting reports highly susceptible to technological assistance.[1]

The researchers then used GPT-4o and Gemini Flash 1.5 as impartial arbiters to reconcile the survey averages with the expert revisions. The models were instructed to evaluate the justifications from both groups, representing "practical knowledge of the work" versus "theoretical knowledge of technology."[1]

The two independent models achieved a 0.96 correlation in their adjusted scores. For example, the task of applying various welding methods received a final score of 0.05, as the models agreed that physical execution remains beyond the capabilities of language models.[1]

In contrast, the task of performing financial estimates to develop a household budget received a final score of 0.64. The models noted that structured budget calculations align perfectly with the data-processing strengths of current generative systems.[1]

Tasks requiring a blend of cognitive planning and physical interaction landed in the middle. Organizing thematic activities for children received a 0.30, acknowledging that while software can suggest ideas, human oversight remains necessary for group dynamics and safety.[1]

Scaling The Predictions

Using the 2,861 adjusted scores as a knowledge repository, the researchers developed an AI assistant to predict values for the remaining 26,892 tasks. This hybrid approach bridged expert-validated human assessments with the scalability required to score an entire national classification system.[1]

The team then applied this predictive model back to the 3,265 tasks in the international classification system used in their 2023 study. The correlation between predictions based on the international tasks and the Polish six-digit tasks reached 0.92, demonstrating high stability.[1]

To ensure precision, the researchers manually reviewed 364 tasks where the 2025 score deviated by more than 25 percent from the 2023 baseline. For 64 highly contested tasks, the authors conducted detailed group discussions to reach a consensus justification.[1]

The final 2025 task-level scores averaged 0.29 across all occupations, slightly lower than the 0.30 average recorded in 2023. However, the score dispersion within occupations narrowed considerably, resulting in a more concentrated distribution of automation risk.[1]

The Four Exposure Gradients

The revised framework abandons the binary "automation versus augmentation" model in favor of four progressive exposure gradients. Gradient 1 represents occupations with low overall exposure but high variability, meaning a few specific tasks might be automated while the core role remains human.[1]

Gradient 2 captures moderate exposure with high task variability, leading to uneven impacts across the occupation. Gradient 3 indicates significant exposure, where a growing portion of tasks face automation risks, necessitating adaptation strategies and retraining for affected workers.[1]

Gradient 4 represents the highest exposure with low task variability. In these occupations, the vast majority of current tasks possess a high potential for automation, and the impact is consistent across the entire role rather than isolated to specific duties.[1]

Clerical occupations continue to dominate Gradient 4. The 13 occupations classified in this highest tier include data entry clerks, typists, word processing operators, accounting clerks, and general office administrators, reflecting the vulnerability of routine data management.[1]

Despite remaining in the highest risk category, the mean scores for several clerical tasks declined compared to 2023. Two years of practical experimentation revealed that tasks like scheduling appointments still require substantial human effort to manage edge cases and context.[1]

The Digital Paradox

Conversely, several highly digitized occupations saw their exposure scores increase between 2023 and 2025. Web and media developers, statistical specialists, and software-related roles moved higher up the gradients as generative models acquired multimodal and agentic capabilities.[1]

In 2023, language models were primarily viewed as text generators. By 2025, these systems could process images, write executable code, and interact autonomously with software environments, significantly broadening the scope of technical tasks they can perform.[1]

While software developers fall into Gradients 3 and 4, the researchers note these roles possess a strong capacity to evolve. Just as bank tellers shifted to advisory roles after the introduction of ATMs, technical workers are expected to develop new tasks using AI tools.[1]

The critical policy question is whether these highly exposed digital occupations will retain and retrain their existing workforce. The alternative is labor churn, where companies replace experienced workers with younger specialists trained natively on the new toolsets.[1]

Global Employment Estimates

Applying these gradients to global labor force data reveals that approximately one in four workers globally holds a job with some level of generative AI exposure. However, the distribution of this exposure is heavily skewed by national income and gender.[1]

In low-income countries, only 11 percent of total employment falls into any of the four exposure gradients. The lack of digital infrastructure and the higher prevalence of manual and agricultural labor naturally insulate these economies from immediate software-driven automation.[1]

In high-income countries, the exposure rate triples. Exactly 34 percent of total employment in advanced economies falls within the four gradients, with 17.3 percent concentrated in the high-risk Gradients 3 and 4, reflecting the dominance of knowledge work.[1]

The gender disparities within these figures are stark. Globally, 5.7 percent of female employment sits in Gradient 3, and 4.7 percent falls into Gradient 4, compared to just 3.1 percent and 2.4 percent for male workers, respectively.[1]

This gap is driven by occupational segregation, as women are disproportionately represented in the clerical and administrative roles that make up Gradient 4. Men dominate the manual, craft, and technical trades that experts scored lowest for automation potential.[1]

The Income Multiplier Effect

The data confirms that as national income rises, the concentration of women in highly exposed clerical roles intensifies. In high-income countries, 9.6 percent of female employment falls into the highest exposure gradient, compared to just 3.5 percent of male employment.[1]

This structural disparity means transition policies must account for the gendered nature of administrative automation. The researchers emphasize that linking this refined index with national microdata provides a foundation for targeted social dialogue and retraining initiatives.[1]

The next verifiable checkpoint will be the integration of these exposure gradients into national labor force surveys over the 2026 reporting cycle. Until statistical agencies track actual task displacement alongside these theoretical exposure rates, the exact timeline of occupational transformation remains an open variable.[1]