What on earth do scientists do with Claude? Anthropic took apart 207,000 scientific research-related conversations from August to September 2026 and looked at it.
Ranking first is data analysis, which accounts for 33.7% in Claude and 48.1% in Claude Code. Writing is a close second at 32% for Claude.
This result is not unexpected, but the ratio illustrates one thing: most scientific researchers use AI as an assistant to handle the most time-consuming "manual work" in the research process, such as running data, sorting results, and writing papers. Core links such as proposing hypotheses and designing experiments account for a much smaller proportion.
There are also differences in subject distribution. Formal sciences such as computers and mathematics are used more per capita; experimental subjects such as chemistry and medicine are used less. The reason is very intuitive: most of the work of the former is done on the computer, while the latter spends a lot of time in front of the experimental bench.
The report also gives an estimate: if existing capabilities are widely adopted, research output in some fields can be doubled; assuming AI takes over all non-manual work, fields such as computers can reach up to 15 times.
Just listen to the number 15 times. Its premise is to "take over all non-manual work," which is far from reality. On the other hand, the judgment of "doubling output" is not an exaggeration when combined with the usage structure above.