Anthropic has published a study of 207,000 Claude and Claude Code research sessions, offering a closer look at how scientists currently use general-purpose AI. The November 5 report finds a concentration in data analysis and writing, with much less use in work tied to physical experiments.
The research announcement says formal sciences such as computer science, mathematics and statistics use Claude more heavily relative to their researcher populations. Data analysis accounts for roughly one-third of the studied Claude sessions and nearly half of Claude Code sessions. Scientists use the systems less often for gathering data and generating ideas, despite valuing those activities highly.
Usage evidence and future scenarios
The full report describes an approach based on automated classification of research interactions. Researchers analyze retained categories and aggregate statistics rather than open-ended details of scientific work. The report also includes a survey of scientists who actively use AI, a population that should not be confused with all researchers.
Alongside observations, the authors model possible increases in research output under different assumptions about adoption and future capabilities. Those estimates are scenarios. The study does not experimentally demonstrate that deploying an assistant causes the modeled increase in discoveries.
What research organizations can learn
Our assessment is that the immediate value lies in identifying where a workflow is already digital and where it remains constrained by the world outside the computer. Writing analysis code can be accelerated without changing the time needed to grow a sample, book an instrument or collect a field measurement. Faster work at one stage may simply move a bottleneck elsewhere.
A lab evaluating an assistant should therefore measure the whole sequence. Useful observations include the time from raw data to a checked result, the amount of researcher correction, reproducibility of the analysis and whether conclusions survive review. Counting generated code or completed conversations would miss much of that picture.
The gap between current use and desired help is also informative. A scientist may readily delegate formatting or scripting while retaining responsibility for choosing an experiment. That can reflect model limitations, access restrictions, difficulty evaluating suggestions or the importance of tacit domain knowledge. Organizations should investigate which explanation applies before assuming that more model capability alone will change adoption.
For software builders, the report suggests that scientific assistance needs support for evidence, provenance and iterative revision. An answer should remain connected to the data and assumptions that produced it, so another researcher can check the result later.
The study provides a substantial snapshot of one vendor’s research usage. Its broader contribution is a framework for asking a more useful question than whether scientists use AI: which stages they trust it with, and what still limits the progress of the overall project.