An Early Look at the Labor Market Impact Potential of Large Language Models

The paper discusses the recent progress in the field of large language models (LLMs) and generative AI. The authors clarify that LLMs can be trained using different architectures and are not limited to transformer-based models. They mention that LLMs can process and generate various forms of sequential data beyond natural language. The authors explain that they use the terms LLMs and GPTs interchangeably in the paper, referring to models like the ones available via ChatGPT. They highlight the importance of integrating LLMs with larger systems to maximize their impact and acknowledge that LLMs can be utilized for tasks other than text generation, such as search applications, summarization, and classification. To assess the potential effects of LLMs on jobs, the authors propose a new rubric that measures the exposure of tasks to LLMs. They collected human annotations and used GPT-4 as a classifier to apply this rubric to occupational data in the U.S. economy. The analysis indicates that approximately 19% of jobs have significant exposure to LLMs, and this number could increase to 49% when considering other generative models and complementary technologies. The authors also examine the correlation between LLM exposure and different factors such as skillsets, barriers to entry, and industry. They find that certain occupations, particularly those relying on programming and writing skills, have higher exposure to LLMs. The paper concludes by stating that the impacts of LLMs are likely to be pervasive and that their potential is expected to increase even without further development. The authors emphasize the general-purpose potential of LLMs and mention that their research provides measurements of LLM impact potential and demonstrates the efficient use of LLMs in developing such measurements at scale. They acknowledge that the evolving impact potential of LLMs over time may be difficult to predict and regulate for policymakers.