The research organization has mapped over 60 distinct tasks in AI development to measure how quickly the field is automating itself
Here’s a question that should keep every AI researcher, investor, and policymaker up at night: how close is AI to automating the process of building AI? Epoch AI, a research organization focused on AI forecasting and trends, just dropped a framework that tries to answer that question with something resembling scientific rigor.
The new taxonomy identifies over 60 individual tasks involved in AI research and development, then maps them against current automation capabilities.
Borrowing from labor economics #
The framework draws directly from ONET, the US Department of Labor’s occupational information network. ONET is essentially the government’s exhaustive catalog of what every job in America actually involves, broken down into granular tasks, skills, and work activities.
This isn’t the group’s first pass at the problem. Epoch AI published findings in 2024 that segmented AI R&D activities into six major categories: hypothesis creation, experiment design, execution, analysis, communication, and studying prior work. The new taxonomy takes that earlier structure and drills much deeper, breaking those broad categories into the 60-plus discrete tasks that make up the daily reality of AI research.
The 2024 work also flagged coding and debugging tasks as the areas with the most immediate automation potential.
Why a taxonomy matters more than it sounds #
The organization has also been building complementary tools. As of early 2026, Epoch AI’s work includes the GATE macroeconomic model, which evaluates AI’s automation impacts on the broader economy. Previous publications from the group have used O*NET data to analyze both remote work automation potential and specific R&D occupations.
David Owen, a noted contributor at Epoch AI, has been among the researchers conducting comprehensive analyses of AI R&D workflows using O*NET-style task labeling, spanning both general automation trends and the specific context of research and development.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our