Generative AI & Work: Automation predictions and dynamic models to identify adverse impacts

The recent release of advanced generative AI models, such as ChatGPT, has reignited concerns about AI-driven automation in the U.S. workforce.

Predicting the impact of generative AI on the labor market is challenging due to its rapidly evolving capabilities and heterogeneous impact across a dynamic economy.

Notably, GPT-4 has surpassed its predecessor in 17 of 26 standardized tests, and generative models often develop unexpected emergent properties as they scale up.

An early study by Eloundou, Manning, Mishkin, and Rock (2023) suggests that up to 56% of worker tasks could be automated by Large Language Model (LLM)-enabled technology, with significant variability across occupations. This technological shift presents both risks and opportunities.

To effectively manage the impact of this emerging technology, there is a critical need for robust, up-to-date predictions on occupational automation and models that can capture the barriers to mobility and dynamics of the labor market.

My research aims to: 1) devise a robust, automated method for predicting occupational automation due to generative AI,and 2) utilize labor flow models to assess the heterogeneous impact of generative AI across the workforce.

My strategy for developing predictions on occupational automation leverages the recent findings of Frank, Ahn, and Moro (2023).

Their research indicates that ensemble methods are more effective at forecasting automation-related unemployment than predictions from individual models.

Additionally, current projections regarding the impact of generative AI have utilized large language models, such as GPT-4, to judge the potential for automating specific tasks. These are then aggregated to the occupational level.

An ensemble of predictive models that can be automatically updated using LLMs, in place of human annotators, would provide robust, up-to-date occupational automation predictions.

Once established, these predictions can be used to simulate an automation `shock' using labor flow models, following the methodology of del Rio-Chanona et al. (2021).

They applied occupational automation predictions from Frey and Osborne (2017) to identify surprising impacts of automation due to complex labor dynamics.

Merging these methodologies promises to identify the parts of the economy that are most vulnerable to the disruption of generative AI automation.