Search

Despite initial research about the biases and perceptions of Large Language Models (LLMs), we lack evidence on how LLMs evaluate occupations, especially in comparison to human evaluators. In this paper, we present a systematic comparison of occupational evaluations by GPT-4 with those from an in-depth, high-quality and recent human respondents survey in the United Kingdom. Covering the full ISCO-08 occupational landscape, with 580 occupations and two distinct metrics (prestige and social value), our findings indicate that GPT-4 and human scores are highly correlated across all ISCO-08 major groups. In absolute terms, GPT-4 scores are more generous than those of the human respondents. At the same time, GPT-4 substantially underor overestimates the occupational prestige and social value of many occupations, particularly for emerging digital and stigmatized occupations. Our analyses show both the potentials and risks of using LLM-generated data for sociological and occupational research. Potentials include LLMs’ efficiency, cost effectiveness, speed, and accuracy in capturing general tendencies. By contrast, there are risks of bias, contextual misalignment, and downstream issues, for example when problematic and opaque occupational evaluations of LLMs may feed back into working life, thus leading to potentially problematic technological constructions of society. We also discuss the policy implications of our findings for the integration of LLM tools into the world of work. [ABSTRACT FROM AUTHOR]