Anthropic has published a study of how scientists use its Claude systems, analyzing 207,049 research-related sessions to identify the fields and tasks where the tools are most frequently adopted. The company found greater relative use in computer science, mathematics and statistics, while health and laboratory-based physical sciences used Claude less heavily compared with their researcher populations.
The main sample covers sessions from August 2 through September 30, 2026 on consumer plans. It includes 128,914 sessions from Claude's Chat and Cowork interfaces and 78,135 from Claude Code, the company's agentic coding tool. Anthropic also examined a separate set of Claude Science sessions, although the supplied report excerpt does not give the size of that supplemental sample.
Researchers used Claude more often for data analysis and writing than for gathering data or forming ideas and hypotheses, the study found. Anthropic says that pattern aligns with an accompanying survey of scientists. It also developed a measure of theoretical AI capability and observed higher use in work rated as more suitable for the technology, particularly cognitive tasks where repeated attempts can receive clear feedback.
To categorize the sessions, Anthropic used language models as classifiers. One classifier determined whether a conversation involved scientific work, while another assigned it to a field, subfield and task in a taxonomy called SciNet. That taxonomy contains 7,262 tasks across 34 fields and 318 subfields, ranging from common activities such as drafting manuscripts to specialized work within individual disciplines.
Anthropic says its privacy-preserving process retained predefined category labels rather than open-ended details about research content. According to the report, classifiers did not store research datasets, hypotheses, experimental methods, early findings or unpublished papers, and the company reported results only in aggregate for groups meeting a minimum number of distinct accounts.
The analysis also models possible effects on research output, but its projections should not be read as measured productivity gains. Anthropic estimates that present capabilities could, in some fields, support output increases as large as twofold through task-level assistance. A much more speculative scenario in which AI automates all non-physical work produces estimates as high as a 15-fold increase for highly cognitive fields. Those figures depend on assumptions about capability and adoption rather than observed scientific breakthroughs.
The authors also identify risks. Reduced human oversight could enable dangerous or unethical research, while automation may change careers, incentives and researchers' sense of purpose. The study's scope creates other limits: it measures use of Anthropic products, not all scientific AI activity, and session classification cannot by itself show whether the work was correct or improved final research quality.
The findings nevertheless offer a detailed picture of current adoption. They indicate that researchers are concentrating Claude use in text- and computation-heavy activities, while physical experimentation and early-stage scientific judgment remain less represented.
