Generative AI and Jobs: A Refined Global Index of Occupational Exposure
Paweł Gmyrek et al. (ILO) · 2025
국제노동기구(ILO)가 생성형 AI가 일자리에 미치는 영향을 직업별로 다시 계산한 2025년 보고서입니다. 세계 노동자 네 명 가운데 한 명이 생성형 AI의 영향을 받는 직업에서 일하고, 사무직이 가장 크게 영향을 받으며, 고소득 국가와 여성의 일자리에서 노출이 더 큽니다. 대부분의 직업에는 사람이 해야 하는 일이 있어 일자리가 없어지기보다 일의 내용이 바뀔 가능성이 크다고 결론짓습니다.
출처 Gmyrek et al. (2025), ILO Working Paper, 140 · 라이선스 CC BY 4.0 · 원문
이 장은 원문만 있습니다. 문장별 해석은 앞 장부터 차례로 붙입니다.
7장 · Adjustment of survey sub-sample
In the next step, we compared the average of expert scores with the original mean survey scores using two independent AI models as impartial arbiters to propose an adjusted score.13 The models were informed about the sources of the scores in such a way that would prevent them from automatically siding with the expert group: group one was defined as “practical knowledge of the work and types of tasks they were presented, since they worked in the same occupational groups from which we selected tasks they scored”, while the second group was described as “respondents with more theoretical knowledge of technology and labour markets”. Rather than simply averaging the two scores, the models were explicitly instructed to focus their analysis on the task characteristics and the current capabilities of generative AI and required to provide justifications for each adjustment and its relative position to the original scores.
The overall correlation between the adjusted scores produced by the two AI models is 0.96, with correlations within occupational groups exceeding 0.9 in all but one case14 (Figure 14). This high level of agreement demonstrates that, when provided with sufficient background knowledge in the form of reference ranges of scores, the models converge on remarkably similar predictions. Notably, only three tasks exhibited an absolute difference in scores exceeding 0.2 (orange dots in Figure 13). In these cases, GPT-4o tended to adopt a more optimistic perspective on the feasibility of automation, whereas Gemini emphasized the importance of human judgment as a critical factor.
For the final adjustment, we averaged the scores of the two arbitrating models and manually reviewed the justifications, without observing any significant inconsistencies. Table 4 illustrates the result of the adjustment process through examples of tasks where expert scores were significantly higher or lower than the survey averages, with the final adjusted score and GPT-4o justification. For tasks such as “operating injection molding machines” or “performing mechanical processing of silicon”, experts highlighted the limitations of generative AI in handling physical and manual components, leading to significantly reduced adjusted scores compared to the survey. Conversely, for tasks like “maintaining documentation of classes with children” or “performing financial estimates for household budgets”, the experts recognized the structured and analytical nature of these tasks, which generative AI can assist with to a substantial degree.
이 글은 위 출처의 원문을 그 라이선스(CC BY 4.0)에 따라 옮겼습니다. 참고문헌 목록과 인용 번호, 그림 · 표 · 각주 · 수식 · 부록은 빼고 절별로 나누어 실었으며, 그 밖의 문장은 원문 그대로입니다. 원 저작자와 발행 기관은 이 화면을 만들거나 검토하지 않았고, 이곳의 해석과 낱말 풀이는 공식 번역이 아닙니다.
This is an adaptation of a copyrighted work of the International Labour Organization (ILO). This adaptation has not been prepared, reviewed or endorsed by the ILO and should not be considered an official ILO adaptation. The ILO disclaims all responsibility for its content and accuracy. Responsibility rests solely with the author(s) of the adaptation.