How 700M+ People Are Really Using AI: OpenAI vs Anthropic

In September 2025, OpenAI and Anthropic released major research on how people actually use AI in the real world.

  • OpenAI's study was published as an NBER working paper analyzing a representative sample of ChatGPT conversations. ChatGPT had more than 700 million weekly users at the time; that platform-wide population is not the study's sample size. It's one of the first large-scale looks at what people actually do with a general-purpose chatbot.

  • Anthropic's September 2025 AI Economic Index (third update) analyzes millions of Claude conversations, links them to the O*NET task database, and adds new views on country/region differences and enterprise API use.

Together, they offer a rich snapshot of how AI is landing in both everyday life and workflows.

1. Different lenses on AI usage

Anthropic focuses on work: it maps Claude conversations to economic activities and asks whether AI is augmenting human work or automating tasks.

OpenAI looks across all use cases: work, learning, personal tasks, curiosity, and more, capturing how AI shows up in everyday life beyond the workplace.

In short:

  • Anthropic → zooms in on workflows

  • OpenAI → zooms out to daily life

2. From "work tool" to "life assistant"

OpenAI's data shows a clear shift over time:

  • Early users were much more likely to use ChatGPT for work.

  • In the study period, a large majority of consumer usage is non-work (roughly 70–73% of messages), with only about 27–30% clearly work-related.

This suggests that AI is evolving from a professional tool into an everyday assistant used for:

  • learning and tutoring,

  • writing and translation,

  • life admin and practical guidance.

Newer users are even less likely than early adopters to use ChatGPT primarily for work, reinforcing the trend toward mass consumer adoption.

3. Automation vs augmentation at work

Anthropic highlights two core modes of AI use:

  • Augmentation – AI as a co-pilot that helps people think, draft, or iterate.

  • Automation – AI independently completes tasks with limited human oversight.

The September 2025 findings suggest that automation-heavy use cases are growing, enabled by improvements in model reasoning and accuracy. At the same time, augmentation remains critical for complex, collaborative, and high-stakes tasks.

Industry patterns:

  • Usage is rising in research-heavy fields like science, education, and math.

  • Some business analysis and consulting use cases show relative decline, possibly reflecting Claude's strengths in reasoning, coding, and technical work.

4. The geography of AI adoption

The Anthropic index also points to clear geographic and economic divides:

  • Richer countries and tech hubs show higher AI adoption.

  • Within the US, states with higher education levels and more tech workers (e.g., Washington, Utah, California, North Carolina) are ahead of the national average.

This mirrors long-standing findings in economic geography: advanced regions tend to adopt new technologies faster, which can widen existing inequalities if not addressed.

5. What enterprises actually do with AI APIs

Anthropic's inclusion of enterprise API usage offers a rare window into company behavior:

  • Around 44% of calls are for programming tasks (code generation, refactoring, debugging).

  • Around 15% are tied to office automation (workflow optimization, documentation, routine tasks).

Compared with consumer usage, enterprises lean more heavily toward automation, driven by efficiency, cost savings, and scalability.

6. OpenAI's behavioral insights

OpenAI's paper and related blog highlights a few broad patterns in ChatGPT interactions:

  • The dominant interaction type is asking questions (information, advice, explanation).

  • The next major type is doing tasks (writing, coding, editing).

  • Purely emotional or expressive conversations are a small minority.

Age and cohort also matter: younger users use ChatGPT heavily but not always for formal work, while some older users are more likely to connect it directly to their job tasks.

7. Why this matters

Taken together, these studies show that:

  • AI has moved well beyond niche expert tooling into mainstream consumer use.

  • Adoption is uneven across regions, sectors, and user groups, which could reshape existing inequalities.

  • Enterprise and individual behavior are diverging: companies focus on automation and efficiency, while individuals often look for augmentation, learning, and guidance.

As more local ecosystems (including China's) build their own large language model infrastructure, behavior-level studies on local users will be critical. They can help policymakers, companies, and educators understand:

  • who benefits from AI,

  • who is being left behind,

  • and where we need targeted interventions to avoid a new "AI divide."