Anthropic researchers published a study examining how scientists use Claude, Claude Code, and Claude Science across various disciplines. According to the report, adoption is highest in formal sciences like computer science, mathematics, and statistics, while health and lab-based physical sciences show lower usage relative to their total researcher population. Across specific research workflows, scientists rely on Claude primarily for analyzing data and writing, while using it less frequently for gathering physical data or generating initial research hypotheses.
The authors attribute these adoption patterns to current AI capabilities, noting that usage correlates with tasks best suited for cognitive reasoning and reinforcement learning. Based on these findings, the report outlines potential acceleration scenarios, projecting that current AI tools could double research output in certain fields through task-level augmentation. In a speculative future scenario where non-physical work is entirely automated, high-cognition fields could experience up to a 15-fold increase in output, creating a feedback loop that accelerates AI development itself.
Why it matters
Identifies computer science, math, and data analysis as immediate high-ROI targets for AI-assisted research tools.
Highlights data gathering and hypothesis generation as key bottlenecks where current LLMs still struggle in scientific workflows.
Suggests AI self-improvement loops will accelerate fastest in computational fields compared to lab-based physical sciences.
Source: anthropic.com



