Anthropic maps how scientists use Claude across 207,000 research sessions
Researchers mostly use the model for data analysis and writing, while lab-heavy fields remain constrained by physical work.
Anthropic has published a large-scale study of how scientists use Claude, based on 207,049 research sessions from consumer plans. The sample covered 128,914 sessions in Claude chat and Cowork and 78,135 in Claude Code, with activity classified by scientific field and research task using privacy-preserving methods.
The clearest pattern is that researchers use AI where the work is already computational. Computer science, mathematics and statistics show the highest relative adoption, while medicine, chemistry and other lab-heavy fields lag because many tasks require physical experiments or direct work with patients. Computer science alone accounted for 32.7% of Claude Code sessions despite representing 4.3% of researchers globally.
Data analysis made up about one-third of general Claude research sessions and nearly half of Claude Code sessions. Writing was also common. Scientists used the model less often for gathering data or generating hypotheses, even though an accompanying survey ranked those activities among the most important parts of discovery.
Anthropic also modeled possible productivity effects. With broad adoption of current capabilities, research output could double in some fields, the company estimates. A more speculative scenario in which AI performs all non-physical research work produces much larger gains in computer science and statistics but only three- to fourfold growth in chemistry and biology.
These are scenario estimates rather than observed productivity gains. The data also covers Claude users rather than the scientific community as a whole, so it shows how one AI system is used—not a complete census of AI adoption in research.
Anthropic
