O*NET for AI R&D

ONET for AI R&D is a proposed taxonomy that breaks down AI research and development work into a detailed list of tasks, so progress toward automating that work can be tracked task by task rather than inferred from indirect proxies. Jean-Stanislas Denain, Joe Kwon, and Anson Ho of Epoch AI introduced it in a June 2026 newsletter post, naming it after ONET, the US Department of Labor database that describes roughly 1,000 occupations and the tasks and skills each one requires. The authors argue that existing ways of forecasting AI research automation, such as extrapolating "effective compute" or METR's task-length "time horizon" metric, rely on proxies that are easy to measure but don't describe what AI R&D actually consists of. They also argue O*NET itself is too coarse for this purpose, since its listed tasks, such as "analyze problems to develop solutions involving computer hardware and software," are too broad to track meaningfully.

The taxonomy organizes AI R&D into six top-level categories covering the research cycle: Decide, Design, Build, Run, Analyze, and Communicate. Each category splits into subcategories, and each subcategory lists specific tasks, for a total of more than sixty tasks compiled from literature review and interviews with AI researchers. Every task carries a rating from 0 to 5 describing how much current AI systems automate it, from 0 ("not used, AI adds nothing") to 5 ("autonomous, end to end with little or no human involvement"), with intermediate points for AI that assists, collaborates, or leads a task under human supervision.

The stated purpose is to let researchers track the fraction of AI R&D tasks that are automated over time, interpret benchmark results by identifying which parts of a job an improved score actually covers, and give frontier labs and forecasters a shared vocabulary for describing where AI systems help with research. The authors describe the initial version as provisional. They expect the category list, task descriptions, and automation ratings to be revised as AI capabilities advance and feedback from other researchers accumulates.