Example of domains, sub-domains, and tasks of AI
Source: Mishra et al. 2025
Note: This taxonomy is derived bottom-up from PapersWithCode’s benchmark database, which tracks over 20,000 tasks across domains. While not identical to top-down frameworks such as the NIST AI Risk Management Framework taxonomy (NIST AI 200-1, 2023), it is data-driven and reflects how research communities define, benchmark, and compare capabilities. The NIST taxonomy provides a conceptual structure organized around categories of functionality and risk, whereas our table represents an empirical curation of tasks and benchmarks as they exist “in the wild.” These two approaches are complementary: NIST offers a governance-oriented lens, while this dataset-driven taxonomy highlights the breadth and granularity of capabilities that are actively measured. Differences mainly trace to what the research community actually measures versus the wide lenses of policymakers. We also recognize that the coverage is uneven (some domains richer in benchmarks than others), which points to areas where further refinement and alignment across taxonomies is needed.
