Most AI labor market research has been looking at the wrong question. The dominant frameworks ask which jobs overlap with current AI capabilities. A more consequential question is which jobs AI can learn to do — and new research from the AI Objectives Institute suggests the answer looks very different from what policymakers currently assume.
The distinction matters enormously. Researchers Philip Moreira Tomei and Bouke Klein Teeselink have constructed a Reinforcement Learning Feasibility Index covering all 17,951 tasks across 894 U.S. occupations in the O*NET database. Their core argument is methodological but carries sweeping practical consequences: reinforcement learning, now the dominant paradigm at the frontier of AI development, is structured around task completion and measurable outcomes. This maps far more directly onto how occupational classification systems are built than the language-model-overlap approach that underpins the most-cited existing indices.
The result is a striking divergence. Gas plant operators, railroad conductors, and chemical plant operators score high on RL feasibility despite scoring low on conventional AI exposure measures. These roles involve monitoring, control, and decision-making within environments that can be simulated and verified — precisely the conditions under which reinforcement learning thrives. Conversely, musicians, physicians, and natural sciences managers score high on LLM exposure indices but low on RL feasibility, because their outputs are subjective, their environments are difficult to simulate, and success criteria are hard to specify computationally.
This inversion has direct policy implications. Current AI workforce programs are largely calibrated around protecting writers, analysts, and software developers — the occupations that score highest on LLM exposure. But the workers in process control, logistics, and industrial operations who face meaningful RL-driven automation risk may fall entirely outside existing policy frameworks. They are, in a meaningful sense, invisible to the current generation of diagnostic tools.
The researchers provide early empirical grounding for their concerns: a difference-in-differences analysis of U.S. job postings shows that occupations with higher RL exposure scores are already beginning to see relative declines in openings compared to less-exposed roles.
For executives managing industrial or infrastructure-intensive workforces, and for investors evaluating automation risk in portfolio companies, this index represents a more operationally honest map of where the next displacement wave will concentrate — one built around AI’s learning trajectory, not its present state.
Source: Raw/ssrn-6659746.pdf