AnthropicAnalysisAnthropic analyzes 207,000 sessions in empirical study on AI scientific use
What happened.Anthropic published an empirical research study examining 207,049 anonymized sessions across Claude and Claude Code to evaluate how laboratory scientists incorporate artificial intelligence into active research workflows.
Why it matters.AI lab roadmaps often presume models will autonomously formulate scientific hypotheses, but real-world researchers use AI as computational copilots for data transformation. Structuring environments around verifiable data analysis and reproducible execution aligns tooling with actual scientific practice. Founders building vertical scientific software should focus on data pipelining rather than speculative hypothesis generation.
The technical specifications:
- Workload distribution: Data analysis accounted for 33 percent of standard Claude sessions and nearly 50 percent of Claude Code interactions, leading all operational categories.
- Workflow boundaries: Scientists consistently reserved early hypothesis formulation and physical experimental data collection for human teams, relying on models primarily for computational execution.
- Writing assistance: Literature synthesis and manuscript preparation represented 24 percent of interactions, with researchers utilizing AI to summarize interdisciplinary citations.
- Specialized tooling: The findings informed design requirements for Claude Science, an integrated workbench featuring persistent code execution environments and automated provenance tracking.
STORYLINE PROGRESSION & CONTEXT
This report represents a verified evolution of prior newsroom intelligence: