Which Jobs Can AI Actually Learn? A New Index Rewrites the Risk List

Here is a number that should stop you in your tracks: the two leading ways researchers measure AI job risk agree only 15% of the time on the jobs that actually matter. Not 95%. Fifteen. For years, the automation "risk lists" you have seen in headlines have been quietly ranking the wrong occupations near the top — and a new study just showed exactly where they went wrong.

The paper, "What Jobs Can AI Learn? Measuring Exposure by Reinforcement Learning" by Philip Moreira Tomei and Bouke Klein Teeselink (May 2026), introduces something the field has been missing. Instead of asking "what can AI do today?", it asks a sharper question: "what can AI be trained to do?" That single shift flips the risk list upside down — and if your job is on it, you deserve to know which side you land on.

The problem with every AI risk list you've seen

Most AI exposure measures — including the widely cited Eloundou et al. work behind so many news stories — score jobs by how much today's large language models overlap with a job's tasks. That sounds reasonable until you notice the flaw: [Fact] large language models are trained mostly on text prediction, so they look "capable" at anything that reads like writing and "incapable" at everything else.

Tomei and Klein Teeselink argue this conflates current capability with learnability. Modern AI systems increasingly improve not by reading more text, but through reinforcement learning — being rewarded for hitting a verifiable goal, over and over, until they master it. The real question for your career is not whether a chatbot can already do your job. It is whether your job produces the kind of clear, checkable feedback signal that lets an AI system learn to do it. [Claim] Those are very different things.

How the RL Feasibility Index works

The researchers built the RL Feasibility Index, or RLFI, and ran all 17,951 O*NET tasks across 894 occupations through it. [Fact] Each task first passes a physical-feasibility gate: if it requires hands, a body, or presence in the physical world, it scores zero. That single gate zeroes out 40.7% of all tasks — a blunt reminder of how much work still lives outside a data center.

Tasks that survive the gate get scored one to ten on eight dimensions tied to how well reinforcement learning could learn them — chiefly, whether success is clearly verifiable and whether a reward signal exists. The scores rescale to a clean 0-to-100 range. [Fact] The occupations that top the list are unglamorous but revealing: data entry keyers at 71%, correspondence clerks at 69%, and proofreaders at 69%. At the very bottom, scoring flat zero, sit dishwashers, stonemasons, and floor layers — physical trades AI simply cannot reach.

The divergence that changes everything

Here is where the study earns its keep. Overall, the RLFI correlates strongly — 0.88 — with the older Eloundou exposure measure. That looks like agreement. But once you filter to only the digitally feasible jobs, the ones where the real fight over automation happens, the correlation collapses to 0.15. [Fact] In plain terms: the two indices agree on which jobs AI cannot touch, and disagree profoundly on which reachable jobs AI can learn.

That disagreement produces some genuinely counter-intuitive results. Consider three jobs the old measures rate as low-exposure but the RLFI flags as highly learnable: power plant operators, railroad conductors, and aircraft cargo handling supervisors. These are structured, rules-driven, heavily monitored roles where success is unambiguous and every action leaves a verifiable trace — exactly the conditions reinforcement learning thrives on. If you operate a power plant control room, the comfortable assumption that your job is "too specialized for AI" may not survive this data. [Estimate] The task structure that makes these jobs safe from a text-predicting chatbot is precisely what makes them learnable by a reward-driven system. You can see the fuller breakdown on our power plant operators page and railroad conductors page.

Now flip it. Musicians, physicians, and natural sciences managers score high on the old exposure indices but low on RL feasibility. Why? Because their real work resists a clean reward signal. [Claim] There is no simple, verifiable score for a moving performance, a correct diagnosis under ambiguity, or the judgment to steer a research team. An AI can generate a plausible melody or a plausible-sounding diagnosis, but "plausible" is not the same as "verifiably correct," and reinforcement learning needs the latter to improve. The detailed picture for these roles lives on our musicians page and natural sciences managers page.

Who is really in the blast radius

The RLFI also maps neatly onto pay and experience in a way that should worry the middle of the labor market most. [Fact] RL exposure is hump-shaped across the wage distribution — it peaks not at the bottom or the top but in the upper-middle deciles, with the authors estimating that a one-log-point rise in salary is associated with a 12.2-point rise in RL feasibility. By seniority it forms an inverted U, peaking at mid-career. [Claim] Entry-level and executive roles sit relatively safer; the solidly established mid-career professional in a structured, measurable job is the most exposed.

And it is not only theory. A difference-in-differences analysis of job postings after ChatGPT's release found that a one-standard-deviation increase in RL exposure was associated with a 2.9% decline in job openings — though the authors are careful to flag this as only marginally significant (p=0.085). [Estimate] The signal is early and noisy, but it points in the direction the index predicts.

What this means for your career

If you take one thing from this research, make it this: stop asking whether AI can do your job today, and start asking whether your job generates a clean, repeatable, verifiable outcome. That is the property reinforcement learning feeds on. Jobs full of ambiguous judgment, physical presence, or accountability that cannot be reduced to a score are, counter-intuitively, the more defensible ones.

For workers in structured, measurable roles — data entry, records administration, routine monitoring and dispatch — the practical move is to lean into the parts of your work that are ambiguous: exception handling, stakeholder communication, judgment calls that no reward function captures cleanly. For those in creative, clinical, or leadership roles the old lists marked as high-risk, this study is a quiet reassurance, but not a free pass — the boundary of "verifiable" keeps moving.

The deeper lesson is about the lists themselves. When the next viral "these jobs will be automated" chart lands in your feed, ask what it actually measured. If it scored jobs on what AI can do today rather than what AI can be trained to learn, treat its rankings with suspicion. The most important risk may be the one the old measures never saw coming.

Sources

• Philip Moreira Tomei and Bouke Klein Teeselink, "What Jobs Can AI Learn? Measuring Exposure by Reinforcement Learning," arXiv:2605.02598 (May 2026). https://arxiv.org/abs/2605.02598

Update History

• 2026-07-04: Initial publication analyzing the RL Feasibility Index (RLFI) and its divergence from existing LLM-based AI exposure measures across 17,951 O*NET tasks.

---

AI-assisted analysis. This article was drafted with AI assistance and reviewed for accuracy against the cited primary source. Occupation-level risk figures link to our detailed data pages, which draw on Anthropic's labor-market research and ONET task data.*