342 occupations, four independent frameworks, one joined dataset. This page visualizes confrontations between theoretical AI capability, observed usage, and pre-LLM automation baselines.
Scroll to explore source-by-source disagreement
The joined occupation table is the core asset in this repo. Filter it, rank it by any major metric, and inspect how the four source views line up for a single job.
The most consequential occupations are not just highly exposed, they also employ millions of people. This view surfaces large labor pools where consensus exposure is already elevated.
X = consensus exposure, Y = occupation employment on a log scale, bubble size = median pay, color = disagreement.
Frey and Osborne captured a robotics-era view of automation. Karpathy captures LLM-era cognitive exposure. The diagonal is equality; points below it are occupations that looked automatable in 2013 but less exposed to current LLM workflows.
OpenAI beta estimates potential task-time reduction. Anthropic observed exposure tracks real deployment. Distance below the diagonal is the adoption ceiling.
Consensus is average score across available LLM-era sources; disagreement is max minus min. High consensus + low disagreement are robust signals. High disagreement flags methodological blindspots.
Occupations where Frey and Osborne rated low risk (`<0.3`) but Karpathy rates high (`>=0.7`). This isolates jobs newly exposed by LLMs.
Workforce-weighted category averages by source. This highlights where all sources align (or conflict) at sector scale.
Developer-class work is both highly exposed and highly rewarded. The same labor segment appears central to automation and to economic demand.