On November 5 Anthropic published a report titled “AI for scientific discovery: How researchers use Claude”. It rests on 207,000 research sessions across Claude, Claude Code and Claude Science, measured with privacy-preserving methods, plus a survey of researchers. The Economics team signs it, with three people marked as lead authors.
Where Claude actually gets used in science
Relative to the number of researchers in each field, use is very uneven. The formal sciences lead: computer science, mathematics, statistics. Health and the lab-based physical sciences trail.
The tasks are even more lopsided than the fields. Analyzing data and writing make up most of it, with data analysis alone at roughly a third of Claude sessions and nearly half of Claude Code sessions. Generating ideas and hypotheses, or gathering data, happens much less. In the accompanying survey, researchers name exactly those two as the most important tasks for discovery.
The in-house comparison is the interesting part: on Claude Science, idea generation takes about twice the share of use that it does on Claude. A tool built for research shifts what the model gets used for at all.
The capability measure explains the pattern
Anthropic builds a measure of how well AI can, in theory, do each research task, then lays usage over it. The top 20% of tasks on that measure get two to four times as much Claude use as researchers’ time spent on them would predict. The bottom 20%, mostly physical work, get almost none. The same pattern holds across fields.
Put simply: researchers use AI where it works. That sounds obvious, but it is the reason the gap between “gets used” and “would matter most” stays open.
Two scenarios, and one number from the survey
From those findings Anthropic runs two cases through a simple model. If today’s capabilities are widely adopted, research output in some fields could double. If AI could take over all non-physical research work, computer science and statistics could see up to a 15-fold increase, while chemistry and biology would rise three- or fourfold, because the bench sets the pace there.
Then there is what the respondents expect themselves: about 40% of researchers who actively use AI think models will carry out full research projects on their own by 2030.
The feedback loop is in Anthropic’s own report
Anthropic writes that the fields with the largest acceleration potential include the ones that produce AI itself, and that this creates an additional loop: faster AI research, more capable AI, faster research again. The fact that this sentence sits in a report from the company that also names its own RSI threshold in the risk factors of its IPO prospectus does not make it wrong. It just makes it harder to read as marketing.