AI widens global economic gap as productivity gains lag: Anthropic
NEW YORK, UNITED STATES — A landmark economic report from artificial intelligence firm Anthropic reveals a deepening “AI divide.” The data suggests that while AI like Claude is accelerating high-level work, global adoption is strictly tied to national wealth, and technical reliability remains a major bottleneck for total labor transformation.
The data, analyzing over one million conversations from November 2025, indicates AI is automating higher-skilled tasks, potentially decimating many professions while offering significant but uneven productivity gains.
Global AI adoption mirrors economic inequality
The Anthropic AI Usage Index (AUI), which measures Claude usage relative to the working-age population, shows a persistent global concentration directly tied to wealth.
A 1% increase in a nation’s gross domestic product (GDP) per capita is correlated with a 0.7% increase in the country’s consumption of Claude per capita.
Moreover, patterns of use vary by income. In wealthier countries, work and personal use are more prevalent, whereas coursework use is most prevalent in countries with the lowest GDP per capita.
In the United States, however, the adoption process is accelerating across states, and forecasts indicate that the country may achieve per capita parity in usage in 2–5 years, about 10 times faster than the adoption of major technologies in the 20th century.
This comparison underscores the mediating role of current economic systems in international gaps, which could lead to a wider gap in AI diffusion.
As the report notes, “These gaps are stable: we see no evidence that low-use countries are catching up or that high-use countries are pulling away.”
AI speeds complex work, but with limits
Claude is primarily used for tasks that require higher education than the broader economy. The mean predicted education for tasks performed by Claude is 14.4 years, compared to 13.2 years for all tasks in the economy.
This leads to greater “speedup,” or time savings, on complex work; tasks requiring a college degree see a 12-times speedup on Claude.ai, compared to 9 times for high school-level tasks.
However, there is a significant trade-off with task success rates, which decline as complexity increases. In API data, success rates drop from around 60% for sub-hour tasks to about 45% for tasks estimated to take humans five or more hours.
The data suggests that in real-world use, Claude achieves a 50% success rate for tasks lasting about 3.5 hours via API and about 19 hours on the interactive Claude.ai platform, where iterative conversation helps break down complex tasks.
AI job impacts depend on task ‘coverage’
The report notes the “effective AI coverage,” which weights task automation by success rates and the importance of each task within a job. This reveals a nuanced picture: some occupations with modest task coverage experience large potential impacts because AI succeeds in their most time-intensive core work.
For example, data entry keyers and radiologists see high effective coverage because Claude proficiently handles their central duties of data processing and interpreting diagnostic images.
Conversely, removing the higher-education tasks that Claude currently handles would lead to a net deskilling effect for most occupations.
For instance, travel agents would be left with more routine ticketing and payment collection, while technical writers would lose analytical and review tasks.
Some jobs, like property management, would require upskilling as AI handles routine bookkeeping, leaving complex contract negotiations to humans.
Productivity boost constrained by bottlenecks
Anthropic’s analysis suggests current AI could boost U.S. labor productivity growth by 1.8 percentage points annually. However, when adjusted for real-world reliability issues, this estimate falls to approximately 1.0 percentage points.
However, adjusting for task success rates reduces this estimate to a range of 1.2 to 1.0 percentage points per year. Even the lower bound of this range remains economically significant, as a sustained increase of 1.0 percentage point would return U.S. productivity growth to late-1990s levels.
The ultimate productivity impact depends on how tasks within a job relate to one another. If tasks are complements (low substitutability), the gains are bottlenecked on activities AI cannot enhance, reducing the aggregate effect to 0.6 to 0.9 percentage points.
The more concentrated, automation-heavy use of APIs is more sensitive to this constraint, highlighting that productivity gains are not automatic and depend on how work is reorganized around AI capabilities.
“Rather than displacing highly skilled professionals, this could instead reinforce the value of their complementary expertise in understanding AI’s work and assessing its quality,” the report suggests.
Anthropic reports that while AI promises significant productivity gains, its current reliability constraints mean its transformative impact on work will be mediated by existing economic inequalities, demanding a fundamental reorganization of labor around complementary human expertise rather than simple automation.
