Scientists hand Claude their writing, not their ideas: hypothesis work is 4.2% of 200,000 research sessions

Scientists mostly use Claude to analyze data and write, and barely use it for generating ideas and hypotheses, the task they themselves rank as most important.

In a report published Nov. 13, Anthropic classified 207,049 research sessions from consumer plans between Aug. 2 and Sept. 30 (128,914 on Claude, 78,135 on Claude Code) using privacy-preserving methods. Data analysis made up 33.7% of Claude sessions and 48.1% of Claude Code sessions, above the 27.6% of time researchers spend on it. Writing and communication took 32.0% of Claude sessions, double its 16.0% share of research time. Ideation and hypothesis generation was just 4.2% (2.1% on Claude Code), even though 528 surveyed scientists ranked it the single most important task for producing results.

Fields differ sharply. Computer scientists are 4.3% of the world's researchers but account for 32.7% of Claude Code sessions, while lab- and patient-heavy fields such as chemistry and medicine use it less. In a hypothetical scenario where AI automates all non-physical research work, Anthropic estimates output in computer science and statistics could rise up to 15-fold. About 40% of surveyed researchers expect AI to complete full research projects on its own by 2030.

Science-of-science researcher Misha Teplitskiy called it "one surprising finding" that scientists in almost every field use Claude "to write rather than code." Replies split between reading writing as an early, immature stage of AI use and users saying the model still needs explicit direction, down to the statistical model, when designing new experiments.