Anthropic's latest Economic Index report, published this week, offers the most granular picture yet of how AI use is actually evolving across the workforce and across the map. The headline number is reassuring in one sense: geographic adoption is converging. The share of Claude usage going to the top five American states has fallen from 30% to 24% over six months. But the same report contains an important counterpoint. Globally, usage is becoming more concentrated, not less. The top 20 countries by per-capita usage now account for 48% of the total, up from 45% six months ago. Convergence at home; divergence abroad.
The distinction matters because the two patterns have different implications. Within the United States, AI tools are spreading to less tech-heavy states at a meaningful pace, though Anthropic's own modelling suggests it will take five to nine more years before usage is roughly equal per capita across states. That is a long time, and the report is careful not to dress it up as a solved problem. But the direction of travel is encouraging. People who work in sectors and regions that were early AI holdouts are beginning to catch up.
The global picture is harder to square with optimistic narratives about AI as a democratising force. Countries that already had high per-capita AI use are widening their lead relative to the rest of the world. The report does not speculate at length about why, but the candidates are obvious: internet access, English-language fluency, disposable income for subscriptions, and a workforce concentrated in knowledge jobs. These are not new inequalities. AI is simply flowing through the channels that already exist.
On the task side, the report tracks something called augmentation, meaning interactions where Claude acts as a collaborator rather than just an executor. The augmentation rate ticked up slightly in both consumer and API traffic. Conversations in Claude.ai also diversified: the top ten tasks accounted for a smaller share of total usage than in November 2025, which means people are bringing Claude into more unusual and heterogeneous situations. That breadth matters. A tool used for one thing is a productivity aid. A tool woven into many different kinds of work starts to change what that work looks like.
The report also tracks something subtler: what kinds of tasks are growing in share. As usage diversifies, the average conversation involves tasks that command slightly lower wages in the labour market. That could be read as AI democratising access to useful tools for lower-wage work. It could also be read as AI automation concentrating first in the tasks and roles with less bargaining power. Probably both are true at once, which is exactly the kind of uncomfortable ambiguity that a well-run research programme should be surfacing rather than resolving too quickly.
Anthropic has been publishing these index reports since late 2025, and the series is quietly becoming one of the more useful data sources in a field otherwise dominated by vendor-commissioned surveys and hot takes. The methodology is disclosed, the findings are often inconvenient, and the framing resists triumphalism. That is not nothing, given who is funding the research. What the index cannot easily answer is whether the patterns it documents are healthy. Converging usage within a wealthy country is better than nothing. But it is not the same as equitable access to the economic gains that AI makes possible, and the two are not guaranteed to arrive together.