“GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models,” posted to arXiv in March 2023, was written by Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock, with the first three authors at OpenAI. The title is a pun: GPTs (the model family) may be GPTs in the economic sense - general-purpose technologies, like electricity or the computer, that ripple across the whole economy. A later version was published in Science.
The paper estimates “exposure” rather than job loss. Using human and GPT-4 ratings of how well tasks across hundreds of occupations align with language-model capabilities, the authors estimate that around 80 percent of the US workforce could have at least 10 percent of their work tasks affected by the introduction of LLMs, and that roughly 19 percent of workers could see at least 50 percent of their tasks affected. They are careful that “affected” means the task could be done significantly faster at the same quality, not that the worker is replaced.
The authors also distinguish raw model capability from real-world tooling. With access to an LLM alone, they estimate about 15 percent of all US worker tasks could be completed significantly faster; once software and applications built on top of LLMs are included, that share rises to between 47 and 56 percent of tasks. The paper notes that higher-wage occupations tend to show greater exposure, breaking from earlier waves of automation that concentrated on routine manual and clerical work. Its exposure estimates became an input to later macroeconomic work, including Daron Acemoglu’s productivity analysis.