A new classification system for artificial intelligence research and development could change how policymakers and investors think about automation in the field. Epoch AI, a research group that studies AI trends, has proposed an O*NET-style taxonomy designed to track which tasks within AI R&D are most likely to be automated.
The taxonomy borrows its structure from the U.S. Department of Labor's O*NET database, which breaks down thousands of occupations into detailed tasks, skills, and work activities. Epoch AI's version would apply that same granular approach to the work done by AI researchers and engineers—mapping everything from literature reviews and experiment design to model training and evaluation.
Why a taxonomy for AI work
Automation isn't just a threat to factory floors and call centers. AI itself is increasingly being used to speed up parts of the research process. Tools that generate code, run experiments, or even write papers are already in use. But until now, there hasn't been a systematic way to measure how much of AI R&D could be handled by machines.
Epoch AI's proposal aims to fill that gap. By categorizing the specific tasks involved in AI development and assessing their automation potential, the taxonomy could give researchers, companies, and governments a clearer picture of where the field is headed. That information could then inform funding decisions, workforce training programs, and regulatory approaches.
What the taxonomy covers
The taxonomy focuses on the entire lifecycle of AI R&D. That includes conceptual work like formulating research questions, practical work like cleaning datasets and tuning hyperparameters, and evaluation work like benchmarking models. Each task gets scored on how easily it could be automated with current or near-term AI tools.
Early analysis using the taxonomy suggests that certain routine tasks—such as hyperparameter optimization and data preprocessing—already have high automation potential. More creative tasks, like formulating novel hypotheses or designing experiments, appear less automatable for now. But the taxonomy is designed to be updated as AI capabilities evolve.
Policy and investment implications
The timing of the proposal matters. Governments around the world are pouring money into AI research while also worrying about job displacement. A taxonomy that shows which parts of AI work are ripe for automation could help target reskilling efforts and avoid investing in areas that are about to be transformed.
For venture capitalists and corporate R&D labs, the taxonomy offers a way to identify bottlenecks in the research process. If a particular task is both critical and highly automatable, that might be a signal to invest in new tools. If a task is resistant to automation, it suggests a place where human talent will remain essential.
Epoch AI has not yet released a full version of the taxonomy for public use. The group plans to refine the framework through feedback from the AI research community before publishing an official version. That means the details—how tasks are classified, what automation thresholds are used, and how often the taxonomy is updated—are still open to debate.
