Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Initial Algorithmic predictions

As the first step, we conducted a purely algorithmic assessment of potential exposure to automation for each task in the Polish 6-digit system, relying on three advanced LLMs available on the market at the end of 2024: GPT-4, GPT-4o and Gemini Flash 1.5. The method relies on the approach previously used by GBB (2023), with a sequential API call devised to assign a score ranging from 0 to 1 for each individual task, where 0 means the task cannot be automated and 1 means the task could be fully automated with GenAI technology (the task can be conducted without participation of a human operator).4 The LLM is provided with information on the occupation to which the task pertains in ISCO-08 and requested to produce a justification for each predicted score.

The overall distribution of these synthetic automation scores is generally consistent with scores based on ISCO-08 documentation (GBB, 2023; Gmyrek, Winkler and Garganta, 2024) and O*NET (Elondou, 2023; Felten, 2023): the highest exposure to GenAI is typically observed among groups with the largest share of tasks of cognitive nature and it declines with tasks of physical nature, which are more prominent among occupational groups 4-9 (Figure 2). The largest density of high automation scores is observed among clerical workers and is stable across the three LLMs. In other groups, we observe some important differences in the LLMs’ predictions. In particular, Gemini Flash 1.5 seems to provide a bimodal scoring, with most scores falling either below 0.25 or above 0.6. The other two models provide a wider range of scores, but still with important divergences, which makes it difficult to judge which model one should follow.5 Therefore, we do not examine these distributions in greater detail but proceed directly to improvements that could enhance coherence and stability of our predictions, regardless of the LLM applied.

Selection of occupational tasks for human assessment

To improve the precision of our synthetic predictions, we design a human-based evaluation of the potential for task automation. Since scoring of all 29,753 tasks would be too cumbersome for a survey, we focus on a sub-sample of tasks that are representative of the tasks in each professional group at 1-digit ISCO-08 level (Table 1). We employ the standard formula to calculate sample size (see Appendix 2).

Table 2 presents task sample calculations for each 1-digit category. The minimum number of scorings per question is set to 20, based on the general rule used by ILO Statistics for micro-data analysis, where observations with less than 16 respondents are flagged as unreliable. The number of scorings is set to 35 per respondent, based on our pre-testing of the survey form for potential fatigue effects.

Within these initial parameters, each respondent is assigned a random set of tasks from the larger sample of tasks for each broad occupational group. The randomization of the tasks’ sample selection maintains a general pattern of the distribution of scores similar to the scores initially generated with the LLMs (Section 4.1). However, the respondents are not made aware of these prior distributions, instead they receive a wide range of tasks with respect to their automation potential.

Survey design and recruitment

We conducted the human survey using the Computer-Assisted Web Interview (CAWI) method, drawing on a large online research platform in Poland, used for both commercial and scientific studies. This also means that some demographic characteristics are already pre-filtered at survey deployment stage, whereas others are captured through a set of core questions at the start of the survey (Table 3). Specifically, we capture respondents’ gender, age group, size and place of residence, region of residence, highest level of education and their current profession. The professional classification is further facilitated by providing respondents with a short description of ISCO-08 1-digit classifications, alongside examples of occupations that fit into each category.

In addition to demographic information, subsequent core questions are designed to capture respondents' knowledge and experience with GenAI. These questions help contextualize how familiarity and interaction with GenAI may influence assessments of task automation potential. Specifically, respondents are asked how long they have worked in their current profession (Q4), as more experienced workers may have a deeper understanding of their tasks and offer more cautious or realistic assessments. Additionally, they are queried on their familiarity with AI (Q5), with more knowledgeable respondents potentially providing more informed evaluations. We also capture how often respondents use AI tools in their job (Q6), which provides insight into their practical engagement with AI and its integration into their daily tasks.

Finally, in core questions 7-8, we ask how respondents view the broader impact of AI on jobs in their field, helping us understand their perspectives on automation over the next five years. The final question in this section probes respondents’ views on the likelihood that their own job could be partially or fully replaced by AI, which helps assess whether job security concerns might influence their task evaluations. Table 3 below provides a summary of the core questions.

In its main section, the survey moves to the assessment of respondents’ perception of tasks’ automation potential. Each respondent is presented with a random selection of 35 tasks from their occupational group at 1-digit ISCO-08 level, as previously described in the sample selection process (Section 4.2). For each task, respondents are asked to evaluate the susceptibility of the task to automation by generative AI on a scale from 0 to 100, with 0 indicating "not automatable" and 100 indicating "fully automatable".6 In addition to scoring tasks for automation potential, respondents are asked to indicate whether they personally perform a similar task in their current job by ticking a box next to each task. This allows for the comparison of GenAI’s perceived impact on tasks that are only present in respondents’ occupational group, with those directly relevant to the respondents' current work.

Text Box 1. Sample Question for Task Automation Assessment

Below is a list of 35 tasks performed by people in different occupations. Please rate how much you think artificial intelligence can automate each of the tasks. Use the slider to indicate your opinion on a scale from 0 to 100, where 0 means no potential for automation, and 100 means full potential for automation.

Task 1: Developing one's own artistic concept of a role

| 0 - No potential for automation using GenAI | ------------ Slider ------------ | 100 - Full potential for automation using GenAI

| I don't know

Do you perform this or a similar task in your current job?

• Yes

• No

Survey design limitations

The survey is designed to capture insights from individuals currently employed in various occupational categories of the 1-digit ISCO-08 framework. This means that tasks of managers are assessed by individuals working as managers, tasks of professionals are assessed by professionals, etc. The subsequent data analysis explores the differences among individuals based on their attitudes about technology, knowledge of AI and other parameters that are captured in the core questions (Table A3).

The advantage of this approach is that the tasks are assessed by those familiar with the work from their own professional context. On the downside, this survey design means that the sample may not represent the entire working population, and interpretability of opinions on the automation potential is limited to comparisons within each 1-digit group. We also expect an imbalance of understanding and experience with tasks and GenAI across different occupational groups, as managers or professionals are likely to know more about AI than elementary workers, whereas the latter are likely to understand better the nature of elementary workers’ tasks, but might have less familiarity with GenAI tools.

We mitigate these negative effects in several ways. First, to equilibrate the level of understanding among participants, a prompt about current abilities of GenAI is shown in the introduction of the second part of the survey (see Text Box 2 and full survey questionnaire in the Appendix). Second, we run a survey among the ILO, NASK and MOL experts on a smaller sample of tasks, drawn from across all ISCO-08 groups, and compare these results to the results of main survey respondents (Section 5). In cases of drastic discrepancies, we consult with the expert group and manually establish what seems to be the most reasonable score.

Text Box 2. Introductory prompt for survey respondents

We have a short piece of information for you about Generative Artificial Intelligence (GenAI). In our study, we want to distinguish GenAI from artificial intelligence (AI).

AI allows you to make decisions based on data (e.g. what product to offer you next in an online store, recognize car license plates at the entrance to the parking lot and raise the barrier)

GenAI can generate new content, e.g. images, text, speech. The most popular tool is ChatGPT, which can answer questions, write emails, prepare an application to the office, create reports, summarize a specified text (e.g. an article, book).

What can GenAI do?

• Write a text, "pretending" to be a specific person, e.g. an official, poet, financial analyst;

• Prepare translations between languages, e.g. from Polish to English;

• Convert written text into human speech ("read" text in a given language);

• Program - change specified tasks into computer language code, e.g. Python, HTML;

• Generate an image, e.g. a photo, based on a given description, or recognize and describe an image (e.g. what is in the photo).

Survey of Task Automation Potential

Sample properties

The main survey sample consists of 57,400 observations and 1,640 unique respondents. Among these, 105 participants responded with 'Don't know' to most questions, providing less than 50% of the required task rankings. Due to limited engagement, these participants were removed from the analysis as they were deemed insufficiently credible. The excluded group includes a variety of demographic and occupational profiles, which were carefully reviewed to ensure that their removal would not introduce significant bias into the results. As the next step, we excluded 1,167 observations where participants responded with 'Don't know' to the ranking of task automation potential. This ensures the analysis focuses only on responses where participants felt equipped to provide meaningful input.

A detailed examination of “don’t know” answers showed that they occur twice as frequently in manual or technical fields – particularly among Technicians and Associate Professionals (Group 3), Craft and Related Trades Workers (Group 7), Plant and Machine Operators and Assemblers (Group 8), and Elementary Occupations (Group 9) – than in other occupational groups.7 Since the differences in stated exposure and frequency of use of GenAI do not seem to justify this pattern (see section 4.5.2), it suggests that participants may have felt less equipped to assess the automation potential of tasks requiring specialized technical expertise or manual skills. This is illustrated though the three tasks with the largest number of “don’t know” responses (7 each): (i) operating devices (centrifuges and filters) used for separating fermented milk mass into yeast slurry, (ii) recording delivered goods and collected coins (tokens), and (iii) organizing and cleaning the workstation.

Following this exclusion, we retained a substantial sample of 1,535 respondents, contributing a total of 52,558 data points. These data points correspond exclusively to cases where participants provided valid responses to the primary survey question regarding the potential for task automation. Importantly, this filtering process did not reduce the total number of tasks assessed, which remains at the representative sample size of 2,861, as calculated in Table 2.

The survey sample is well-balanced in terms of gender, with 51% female and 49% male respondents. The actual gender distribution among the employed population aged 18+ in Poland is more skewed toward men, with 53.9% of employed individuals being male and 46% female. As shown in Figure 3, this results in the survey overrepresenting female opinions in the younger age brackets (18–29 and 30–39).

Figure 4 illustrates the broader distribution of the survey sample across ISCO-08 1-digit groups, compared to both the employed population in the LFS and the desired sample (Table 2). Professionals (2) are clearly the most underrepresented group, while groups 1, 8 and 9 are overrepresented in survey responses, compared to the actual employment numbers. Notably, however, the desired sample calculated in Table 2, based on the share of tasks in each 1-digit category, achieves a closer alignment to the LFS structure for certain occupational groups.

In terms of sex-related representation within occupations, largest imbalances occur within occupational groups of public authorities, senior officials and managers (1), clerical workers and elementary occupations (9), which are represented by women in the survey to a visibly higher extent than in the labour market (Figure 5). Men are underrepresented, compared to the LFS, among professionals (1) and craft workers (7). For other occupational groups, the survey pattern is fairly consistent with national labour market trends.

Such discrepancies are important to note but do not affect our strategy, as the primary goal of sampling was to ensure representativeness of tasks at the level of occupational 1-digit groups, rather than replicating the exact gender and economic structure of the employed population (Table 2). With this objective met, we did not apply additional weight-based adjustments to the survey sample. Instead, the deviations between the survey and LFS are treated as a natural variation inherent to a task-specific character of this study, which follows the imbalances in the total number of tasks across 1-digit groups, as well as other assumptions made in Section 4.2.

Participants’ Exposure to GenAI: Screening Questions

The screening section of the survey tested participants’ prior exposure to GenAI. A significant majority of survey participants (71.2%) report having heard about this technology, with relatively small differences between genders and occupational groups. Despite this high level of awareness, the actual use of GenAI is limited, with most respondents reporting they "never" (50.1%) or "rarely" (26%) use it. Sporadic use (few times a month) is reported by 17.4% of all participants, with only 4.8% reporting regular use and less than 2% using it very often. Among the 1-digit occupational groups, professionals, technicians, public authorities, senior officials and managers use these technologies more frequently than the groups with a stronger presence of manual tasks (Figure 6). This seems consistent with recent studies of GenAI adoption, which suggest higher use of GenAI among occupations with a larger share of cognitive work (Bick et al., 2024; Parker et al., 2025).

When considering the potential impact of GenAI on jobs within their industries, most respondents expect either no impact (26.5%) or only a slight impact (30.2%). As shown in Figure 7, manual and technical groups have a significantly higher share of such responses. Reports of moderate and significant impact are highest among professionals (33.8% and 20.4%), public authorities, senior official and managers (31.5% and 24%), clerical support workers (22.5% and 21.1%) and technicians (27.3% for moderate impact), though the latter stand out with notably lower expectations of a significant impact (14.4%). This might be due to the inherent complexity of many tasks in this professional group. In comparison, relatively few respondents predict a complete transformation of their occupations, with the highest share of such responses among service and sales workers (5.1%) and clerical support workers (4.2%), and a small total number of these observations.

We further examine these trends by asking respondents a more personal question, related to their perceived likelihood of GenAI replacing their current job in the next five years. For most professions, responses are clustered at the lower end of the risk spectrum, suggesting limited concern about replacement by GenAI in the mid-term (Figure 8). However, among clerical support workers, the density of responses representing moderate to high levels of concern is noticeably higher, with women reporting slightly higher risks compared to men, while the converse is observed among plant and machine operators and assemblers.

Scoring of task automation potential

The survey asked respondents to evaluate the extent to which GenAI could automate the execution of various tasks (see Text Box 1). As shown in Figure 9, responses reveal gendered scoring that aligns with occupational segregation. For example, the scores provided by clerical support workers are predominantly from women, reflecting the high representation of women in this occupational group. However, men within this occupation tend to assign higher scores, potentially reflecting a different perception of task automatability or biases regarding clerical work, despite their lower representation in this group. Conversely, men dominate responses in occupations such as craft and trade workers and plant and machine operators, but for the tasks in these occupations, they assign lower scores than women. These differences may be representative of different task composition or even tasks performed by men and women, even when in the same occupational category, which then influence their scoring.

To better understand the factors associated with higher task automation scores, we conducted a set of exploratory regressions using the main task-level score as the dependent variable and controlling for occupational clustering. Results suggest that across occupations, familiarity with a task is associated with higher scores, and this relationship strengthens with greater reported exposure to GenAI. Respondents who use GenAI frequently – such as on a daily or weekly basis – tend to assign higher scores. Similarly, those who believe GenAI will have a “Complete” or “Significant Impact” on their occupation consistently assign the highest task scores (Figure 10).8

One way to address this variation in scoring would be to apply corrective weights to responses, giving greater influence to respondent profiles likely to have more relevant expertise. For example, one could assign more weight to respondents reporting greater familiarity with GenAI or frequent usage or prioritize scores from individuals who perform similar tasks in their daily work. However, in our sample, this would introduce significant limitations. For 91.5% of the tasks, respondents did not report performing a similar task in their work – even when the task belonged to the same 1-digit occupational group as their current job. This is not surprising. For instance, within the “Professionals” group alone, the representative sample consisted of 368 tasks drawn from a pool of 8,715 tasks (Table 2), making exact task matching unlikely.

Moreover, adjusting for beliefs or usage would require strong theoretical assumptions about what constitutes “credible” scoring. For example, a respondent who believes their profession will be completely transformed by GenAI in the next five years may be highly informed, or may be expressing anxiety, speculation, or dystopian fears. Without richer data to distinguish between these interpretations, applying corrective weights risks introducing new biases.

Given the relatively small size of these specific subgroups in our dataset, we refrain from implementing such adjustments in the current analysis to avoid overfitting or compromising credibility. Instead, we flag this as an important area for future survey design, particularly in capturing respondent expertise and task familiarity more systematically.

For the present study, we treat the survey as a starting point, designed to capture a broad spectrum of perspectives reflecting varying degrees of familiarity with both tasks and GenAI. In many cases, a respondent’s understanding of the task itself may be more influential than their understanding of the technology, especially for concrete physical tasks (e.g. trimming flowers, assembling mechanical parts). Accordingly, we rely on mean scores across all respondents to reflect the collective judgment of a diverse sample grounded in workplace experience. To enhance robustness, we supplement this with expert evaluation of a subset of tasks (section 5), ensuring that final scores reflect both practical insights and a realistic appraisal of current GenAI capabilities.