How do scientists use an AI assistant in the work they actually do? Anthropic published a study on November 5, 2026 that classifies 207,049 research-related Claude sessions from August 2 through September 30. Its answer is more specific than a claim that AI is doing science: data analysis and writing dominate, while collecting data and forming new hypotheses appear less often. The sample describes activity on Anthropic's products, not all researchers or all AI tools.
What did Anthropic measure in the Claude research study?
Anthropic says its main sample contains 128,914 sessions from Claude Chat and Cowork and 78,135 from Claude Code, all on consumer plans. Classifiers assigned a scientific field, subfield and task from a taxonomy of 7,262 tasks. The research team says it retained the category labels rather than open-ended details from the conversations and reported aggregate cells only when enough distinct accounts contributed. It also surveyed 528 scientists, mainly active AI adopters in life and health sciences. A conversation classified as research work does not by itself prove the user was a scientist or that a useful discovery resulted.
Which research tasks are most common?
Data analysis accounts for 33.7% of the sampled Claude research sessions and 48.1% of Claude Code sessions. Writing and communication account for 32.0% on Claude and 17.6% on Claude Code. Data gathering is 11.5% and 11.9% respectively, while ideation and hypothesis generation are 4.2% and 2.1%. These are shares of classified sessions, not shares of scientific discoveries, time saved or a measured productivity lift.
The report compares those shares with estimates of how researchers spend their time. Data gathering takes a much larger portion of research time than of Claude sessions, partly because experiments, patient contact and fieldwork require physical activity. The survey participants ranked ideation and data gathering as especially important for discovery, even though they reported using AI less for those tasks. Anthropic acknowledges that ideation can be hard to classify when it is woven into another task.
Why does Claude Code show more computer science use?
Computer science accounts for 32.7% of Claude Code research sessions in the sample, compared with an estimated 4.3% share of researchers worldwide. That fits a tool built to work with files and run code, and the report finds relatively high use in computational subfields across other disciplines too. It does not mean one-third of computer scientists use Claude Code. The denominator is Anthropic's research sessions, and people using other products or no AI are outside it.
Does this show that AI has doubled scientific output?
No. Anthropic models possible acceleration under assumptions about adoption, task capability and cost. Its estimate that current capabilities could double output in some fields is a scenario, not an observed change in published research. More dramatic projections require greater automation of non-physical work; the report says lab and field work remain bottlenecks and notes that its calculations omit some equipment costs, shared facilities and the quality of ideas. The usage study is useful precisely because it separates observed behavior from these forecasts.
For people building research software, the pattern suggests a concrete design question: can a tool make data analysis or writing easier while preserving the sources, steps and human checks needed to trust the result? VIBPS is the social network for software, including tools made for research, but it has no role in this study and has not independently audited Anthropic's classifications.