Though large language models (LLMs) and associated technologies are poised to disrupt jobs, little has been done to evaluate this risk or develop policy to address it. In a Policy Forum, Tyna Eloundou and colleagues report results of a rubric they developed to assess effects of LLMs on labor markets. Higher-wage occupations are more exposed to LLMs, they report. Despite a growing literature that aims to understand the labor market impacts of LLMs, there is still no clear understanding of how "exposure" to AI systems will translate to real-world impacts on labor demand, wages, inequality, job quality, and other key outcomes. Here, Eloundou and colleagues used the O*NET 27.2 database – which covers dozens of occupations and their detailed workflows – and focused on 923 occupations, in particular. They used humans and trained GPT-4 to evaluate if an LLM could reduce the time required for a human to complete a task by at least 50% while preserving or improving quality. The authors estimate that about 80% of workers are in occupations with at least 10% of job tasks exposed to LLM influence in this way, with 18.5% of workers in occupations with 50% of their tasks exposed. An additional exploratory analysis revealed 1.86% of tasks could be fully automated with no human oversight. Eloundou et al. found that, generally, higher wage occupations were more exposed to the effects of LLMs than lower wage occupations. In fact, the two job groups most exposed to LLMs are "Scientists and Researchers" and "Technologists." However, say the authors, tasks currently considered out of reach of LLMs might become more viable with future innovations, and conversely, tasks that seem more likely to be handled by LLMs might encounter unexpected barriers. This work highlights the importance of developing policies to ease transitions in LLM adoption and direct their development in a more socially beneficial direction, ensuring that economic gains are large and broadly distributed.
For reporters interested in trends, an April 2022 study published in Science Robotics investigated a way to tell how at-risk a job is of being taken over by robots or artificial intelligence. It also proposed strategies to mitigate this outcome.
