This figure from Anthropic comparing “theoretical AI capability” with observed usage across occupations has been circulating widely in the AI policy bubble. The gap between what AI could supposedly do and what people actually use it for speaks for itself.
But two small irritations.
First, the word “theoretical.” Most of these exposure estimates are not theories in any meaningful scientific sense. They are constructed by mapping occupational tasks to model capabilities and asking whether AI could plausibly perform them. That produces something closer to a hypothesis about potential exposure, not a theory.
Readers familiar with the development world may recognize the pattern. Just as most (if not all) “theories of change” would be more accurately described as hypotheses of change, many “theoretical exposure” estimates are essentially structured guesses.
Second, if these exposure estimates are hypotheses, their value should lie in predicting actual usage. The figure itself suggests they do not do that particularly well. Several occupations predicted to have substantial exposure show relatively little activity.
Which raises a simple question: why do we keep treating “theoretical exposure” as the main way of thinking about AI’s labor market impact? At this point we have something better: usage data.
Usage reflects constraints that exposure models miss, such as workflows, incentives, legal responsibilities, organizational risk tolerance, and the simple fact that integrating AI into real work is harder than demonstrating that it could perform a task.
So the interesting lesson from this figure may not be the gap between theory and practice and, instead, that AI adoption follows institutional and contextual feasibility much more than technical capability.