Two significant studies released this week offer a comprehensive view of how AI is actually being used, based on data from millions of interactions. Anthropic’s Economic Index Report and a National Bureau of Economic Research paper co-authored by OpenAI researchers analyse usage patterns from Claude and ChatGPT respectively, revealing a technology spreading at historic speeds yet concentrating in unexpected ways.
The OpenAI study tracked ChatGPT’s consumer growth from launch through July 2025, when it reached 700 million weekly active users sending 2.5 billion daily messages. Using privacy-preserving automated classification of conversations, researchers found that non-work usage has surged from 53% to 73% of all messages in just one year. Nearly 80% of conversations fall into three categories: practical guidance like tutoring, information seeking, and writing tasks. Writing dominates work usage at 40%, though two-thirds involves modifying existing text rather than creating new content.
Coding accounts for only 4.2% of ChatGPT messages, while companionship uses barely reach 2%. The preference seems to be for “asking” over “doing”; 49% of messages seek information or advice rather than task completion. This preference is strongest among educated professionals, suggesting high-skill workers value AI primarily for decision support rather than automation.
Anthropic’s report, analysing both consumer and enterprise API usage through August 2025, tells a different story. Coding dominates at 36% of consumer usage and 44% of enterprise. The study exposes geographic concentration: Singapore uses Claude at 4.6 times its expected rate based on population, while Nigeria sits at just 0.20 times. Within countries, usage correlates strongly with income and education levels.
The enterprise data reveals how businesses actually deploy AI. API usage is 77% automation-focused, with companies systematically delegating complete tasks rather than iterating collaboratively. But Anthropic identifies a barrier that we have discussed recently: complex tasks require disproportionately more contextual input, suggesting that organisational architecture, not model capabilities, limit sophisticated deployment.
The papers agree on AI’s unprecedented adoption speed and its uneven distribution across demographics and geographies. Both identify education and income as key predictors of usage intensity. Yet they diverge sharply on coding prevalence and interaction patterns. Where OpenAI sees rising demand for decision support, “asking”, Anthropic observes increasing automation or “doing”, particularly in the enterprise.
These differences likely reflect platform specialisation and user selection. Claude appears to attract technical users and challenging business use cases, while ChatGPT serves broader consumer needs. The coding discrepancy alone suggests AI tools are fragmenting into specialised niches rather than converging toward universal assistants.
AI is clearly creating two distinct change profiles, a consumer economy focused on daily assistance and learning, and an enterprise economy pursuing aggressive automation. The concentration of sophisticated usage in wealthy regions and among educated professionals suggests AI might accelerate rather than reduce inequality. But at the same time, lower-income countries’ preference for automation over augmentation could mean that they will get the chance to leapfrog the more embedded and complex processes that are harder to automate.
Takeaways: These studies reveal AI adoption following familiar patterns of technological diffusion, just compressed into months rather than decades. The diversity of “non-work” usage suggests we need new metrics for measuring AI’s value creation beyond traditional economic indicators. But with AI’s rapid progress the patterns these papers document could start to crystallise before we understand their implications.