Anthropic research reveals that more than half of roughly 250,000 Claude conversations involve users delegating consequential tasks to artificial intelligence. The findings challenge the assumption that AI is used primarily for low-stakes work.

The study was conducted by three independent groups using Anthropic’s privacy-preserving platform, Anthropic Insights. Researchers from Stanford University’s Social and Language Technologies Lab, the University of Oxford’s Human Information Processing Lab, and AI evaluation nonprofit METR analyzed the data.

Collaboration Over Automation

Stanford researchers examined how people collaborate with AI and how much control humans retain. Users were particularly likely to bring consequential work to Claude when seeking professional guidance, especially on legal and financial questions.

In nearly three-quarters of conversations, people set the direction while Claude assisted. Users typically adapted responses rather than using them verbatim, suggesting AI use is closer to collaboration than full automation.

User Engagement and Friction

Researchers found that interactions often involve friction, which can lead to better results by forcing people to refine their thinking. Users may realize Claude misunderstood a request, clarify their intent, and revise instructions.

Oxford researchers looked at user feelings during interaction. Warmer responses from Claude coincided with more positive reactions, while refusals were associated with users pushing back. More unusual responses appeared to coincide with greater intellectual engagement.

Patterns of absorption, frustration, and enjoyment in Claude conversations looked similar to patterns observed during everyday internet browsing.

Independent Research Access

Anthropic aims to give outside researchers access to real-world AI usage data without compromising privacy. Researchers received aggregated results through Anthropic Insights after privacy and legal reviews. A third-party privacy audit by Imperial College London examined the data-sharing process.

Anthropic acknowledged the system made research slower and more resource-intensive but said the pilot demonstrated independent research is possible. The company is considering expanding the program and has publicly released aggregate data from the three studies.

Anthropic has agreed to pay Nscale $45 billion for AI computing power. The deal underscores the escalating capital requirements for large-scale artificial intelligence infrastructure.

The transaction value of $45 billion represents a substantial financial commitment by Anthropic to secure necessary computational resources from Nscale. No further details regarding payment terms, duration, or specific hardware specifications were disclosed in the source material.

What the Numbers Show

The sheer magnitude of the $45 billion figure highlights the intensifying competition for AI compute capacity. This single data point indicates that infrastructure costs are becoming a primary determinant of scale in the generative AI sector, though no comparative benchmarks or historical context were provided in the source.