AI for scientific discovery: How researchers use Claude

A study of 207,000 Claude and Claude Code sessions examines how researchers use AI across scientific fields and research tasks, and what those patterns imply for scientific productivity. Privacy-preserving classifiers assigned sessions to fields and tasks from a taxonomy covering 34 fields and 7,262 research activities; researchers retained only aggregate labels, not details such as hypotheses, datasets, or findings. Usage relative to the number of researchers was highest in formal sciences such as computer science, mathematics, and statistics, and lower in health and laboratory-based physical sciences. Claude Code use was especially concentrated in quantitative work: computer science accounted for nearly a third of its sessions despite representing a small share of researchers.

Researchers mainly use Claude to analyze data and write, with data analysis accounting for about one-third of general Claude sessions and almost half of Claude Code sessions. They use it less for idea generation and data gathering, although researchers surveyed considered those activities especially important to discovery. Usage aligns with estimated AI capability: tasks AI handles well receive two to four times the usage expected from the time researchers spend on them, while physical tasks receive little. A scenario model suggests widespread use of current capabilities could double output in some fields. If AI performed all non-physical research work, output could rise up to 15-fold in fields such as computer science and statistics, but only three- or fourfold in lab sciences constrained by physical work. The report also notes risks to oversight, research ethics, and researchers’ careers.