Artificial intelligence can replace tasks and contribute to job cuts, but a job being exposed to automation does not mean it will disappear. The index published by the International Labour Organization (ILO) and Poland’s NASK research institute on 20 May 2025 estimates which occupations contain activities that generative AI could technically perform. It does not measure layoffs that have occurred or predict how many workers will lose their jobs in Brazil.
That distinction changes how the study’s most widely quoted figure should be read: about 25% of global employment is in occupations with some degree of exposure. It does not mean a quarter of workers will be dismissed. The researchers consider changes to job tasks the more likely outcome, while acknowledging that risks differ across occupations and adoption conditions.
What the ILO and NASK actually measured
The starting unit is the task, not the entire job. An occupation combines different activities: drafting a document, checking information, helping a person and making decisions may have very different automation potential. The index estimates the extent to which such tasks could be performed using generative AI, technology that can produce text, images and other content in response to instructions.
In Working Paper 140, researchers used a Polish occupational classification containing 29,753 tasks. A survey of 1,640 workers yielded 52,558 assessments of the automation potential of 2,861 selected tasks. Their responses were checked against expert judgments; AI models then helped extend the scores to the International Standard Classification of Occupations (ISCO-08). These figures describe how the method was built, not 29,753 lost jobs or dismissed workers.
Each task receives an automation-potential score from 0 to 1. To classify an occupation, the ILO considers both the average score and its spread, meaning whether the activities have similar or very different levels of exposure. The occupation scores are then matched with labour-force surveys from more than 140 countries. This estimates technical exposure rather than directly observing companies deploying AI systems.
The four exposure gradients explained
The index distinguishes occupations that are unexposed or minimally exposed and, among exposed occupations, uses four increasing gradients. A high gradient indicates more widespread technical potential across a job’s tasks, not an individual probability of being laid off.
| Gradient | How to interpret it |
|---|---|
| 1: low | The occupation still depends mostly on tasks with low automation potential, although some activities are more exposed. |
| 2: moderate | Highly exposed tasks coexist with less affected ones; change may be uneven within the same job. |
| 3: significant | A substantial share of activities has automation potential, although less exposed tasks remain. |
| 4: highest | Potential exposure is high and relatively consistent across tasks. Even here, the entire job cannot be assumed to disappear. |
In the ILO classification, clerical roles are among the most exposed. Digital work in areas such as software, finance and media also appears, but there is no single verdict for everyone in a sector: specific tasks and how work is organized matter.
What the global 25% and 3.3% figures mean
The ILO estimates that one in four workers worldwide is in an occupation falling within gradients 1 to 4. Only 3.3% of global employment is in gradient 4, the highest-exposure group. That 3.3% is included in the 25%; it is not an additional percentage of layoffs.
Exposure varies with the composition of the labour market. In the 2025 study, the estimated share of exposed occupations was 34% in high-income countries and 11% in low-income countries. Within global employment, gradient 4 also differed between jobs held by women (4.7%) and those held by men (2.4%). None of these percentages tells us how many people actually lost or will lose their jobs.
What about Brazil? A national exposure estimate is not a layoff forecast
The global 25% figure cannot be multiplied by Brazil’s workforce to forecast dismissals. Assessing exposure in Brazil requires matching occupational scores to the country’s employment structure while accounting for job classifications, the period covered and the statistical source.
An analysis published by FGV IBRE on 7 January 2026 performed that mapping using data from Brazil’s Continuous National Household Sample Survey (PNAD Contínua). For the third quarter of 2025, it classified 29.8 million employed people, about 30% of the employed population, in gradients 1 to 4. The calculation excludes minimal and no exposure. This is an occupational exposure estimate based on matching job classifications with national data, not a count of workers replaced by AI.
The same analysis estimated exposure among 35.4% of employed women and 25.2% of employed men. These results describe how employment is distributed across technical exposure categories; they cannot attribute observed differences in unemployment to AI. Nor do they imply that the Brazilian mapping and the global index have exactly the same occupational mix or reference period.
Why automation capability does not automatically mean layoffs
Between a tool being technically able to carry out a task and a job disappearing lie choices and barriers that the index does not measure. Employers must consider cost, infrastructure access, output quality, data protection, rules of use and human oversight. Even if an activity is automated, time saved may shift to other work, and employment can rise, fall or remain unchanged.
Consider an illustrative example: a system might help draft and organize reports while professionals remain responsible for checking the facts, handling exceptions and being accountable for the result. Some tasks may be highly exposed without the occupation being fully replaced. The reverse is also possible in particular settings: a company may reduce hiring or positions after reorganizing processes. The index alone cannot tell which path will occur.
To establish actual job losses, researchers would need to track employment, hiring, separations, wages and real AI use over time while accounting for other economic factors. The ILO/NASK study maps where change could happen; it does not causally measure what the technology has already caused.
How to read a headline about AI and jobs
- Ask whether the figure concerns potentially automatable tasks, exposed occupations or observed layoffs. They are different measures.
- Check the source, date and geography: a global average is no substitute for research on Brazil’s labour market.
- Look for exposure gradients and whether the analysis includes tasks that still need people.
- For claims that jobs have been eliminated, require labour-market evidence and a method for causal attribution, not merely AI capability scores.
The 2025 evidence identifies occupations where generative AI may change more tasks; FGV IBRE’s Brazilian analysis helps locate that exposure nationally. Neither analysis, on its own, provides a forecast of layoffs in Brazil.
