Anthropic’s research-use study finds AI concentrated in analysis and writing

Anthropic published a study on November 5 examining 207,049 scientific-work sessions on consumer Claude plans between August 2 and September 30, 2026. The sample includes chat, Cowork and Claude Code. It offers a view of where researchers are using one company’s systems, rather than a census of scientists or a direct measurement of discoveries produced. Anthropic study
The report finds heavier use in formal sciences such as mathematics, statistics and computer science, relative to researcher numbers. Data analysis accounts for about a third of Claude sessions and nearly half of Claude Code sessions. Writing is also prominent; idea generation and data collection receive less use. Reported usage patterns
Sessions measure activity, not outcomes
A session count is different from a count of people, papers or successful projects. One researcher can create many sessions, and a difficult task may generate more interaction precisely because it is not going smoothly. Conversely, a brief answer could be highly valuable without producing a long trail of use.
That makes the data useful for identifying adoption patterns but insufficient for concluding that the most-used task is where AI creates the greatest benefit. A productivity claim would need a separate measure of quality, time or completed work, together with a credible comparison.
Anthropic says automated classifiers assign research fields and tasks, with the research team retaining category labels rather than open-ended details about unpublished findings. Results are aggregated with minimum-account thresholds. This provides information about the study’s privacy design while leaving classification accuracy as a methodological question readers should consider. Study methods
Adoption varies within science
Earlier evidence also points to a distinction between trying AI and integrating agents. In a survey published in May, Anthropic reported that 81% of 1,260 social-science respondents had tried chatbots for research, while 20% had adopted coding agents. That survey was fielded in February and March and covers a different population from November’s usage study. Earlier social-science survey
The two sources should not be combined into a trend line. They use different samples and measures. Together, however, they suggest a useful question for research organizations: which parts of a workflow have actually changed, and which still rely on occasional experimentation?
A team evaluating an assistant can answer that locally by separating drafting, analysis, verification and physical work. Record where the system saves effort, where it shifts effort to checking and where it creates a dependency on tools or data that colleagues cannot access.
Scenarios are not measured acceleration
The November report explores hypothetical increases in research output, including much larger gains if all non-physical research work could be automated. Those calculations are scenarios under assumptions, not observed productivity results. Research-acceleration scenarios
For a laboratory or university, the more actionable first step is narrower: choose a recurring task, define an acceptable result and compare assisted work with the existing process. Include review, reproducibility and correction time rather than only the speed of the first answer.
Anthropic’s study helps show where AI is already entering research. Establishing how much it improves science requires following that activity through to reliable outputs, not treating engagement with an assistant as the outcome itself.
Sources and citations
Revision history
Published .