Anthropic’s Science Study Shows Where AI Helps—and Where Labs Still Wait

AI appears most often in the parts of scientific work that already fit inside a browser, notebook or code editor. Anthropic’s new analysis of 207,049 Claude and Claude Code sessions found that data analysis and writing dominate recorded use, while physical data gathering and hypothesis generation appear less. The pattern is revealing, but the measurement is about use of one provider’s tools, not whether those sessions produced better science.

What the report actually measured

Anthropic says its main sample contains 207,049 research sessions on consumer plans between August 2 and September 30, 2026: 128,914 from Claude and 78,135 from Claude Code. A classifier labeled sessions using a taxonomy of 7,262 research tasks across 34 fields and 318 subfields. Anthropic says the research team retained the resulting categories rather than open-ended details such as research datasets, hypotheses or unpublished findings, and reports only aggregate cells above a minimum account count. ^{[1] [2]}

Within that sample, Anthropic reports that data analysis accounted for 33.7% of Claude research sessions and 48.1% of Claude Code sessions. Writing and communication accounted for 32% and 17.6%, respectively. It compares these shares with estimated researcher time and finds that data analysis and writing are more represented in sessions than in researchers’ time, while data gathering and ideation are underrepresented. These percentages are shares of classified sessions, not a share of all global research activity or a measure of time saved. ^[1]

The gap between common tasks and important tasks

The report’s sharpest observation is not simply that scientists use Claude for writing and analysis. In a survey of 528 scientists who actively use AI, respondents ranked idea generation and data gathering as especially important for producing research results, even though those areas appeared less often in tool use. Anthropic offers several possible explanations: physical work is difficult to delegate through a chat interface, while ideation may be woven into other tasks or may still exceed current systems’ strengths. ^[1]

That gap points to a product-design problem. A tool that makes routine analysis faster does not automatically help a laboratory decide which experiment to run, obtain samples, calibrate equipment or judge whether a result is meaningful. Specialized tools may connect to domain databases and make more research tasks accessible, but the work still depends on validation and human expertise. The report’s separate analysis of Claude Science suggests the task mix can change in a specialized workbench; it does not establish that those workflows result in more confirmed discoveries.

The broader research literature also studies AI use through surveys and publication records, each with different sampling and measurement choices. A German researcher survey asks who uses AI and for what, while a Nature analysis examines publication data and researcher impact. Those studies provide context, but they do not independently validate Anthropic’s classified-session counts because they measure different populations and outcomes. ^{[3] [4]}

A scenario is not an observed productivity result

Anthropic’s report models scenarios in which existing capabilities could double research output in some fields, and a hypothetical case where AI performs all non-physical research work could raise modeled output much further in computer science or statistics. Those are scenario calculations, not a controlled comparison of teams that used AI against teams that did not. They depend on assumptions about capabilities, adoption and how research tasks fit together. ^[1]

Readers should therefore treat the report as a descriptive signal about current product use and as a set of hypotheses for further evaluation. The strongest follow-up evidence would track completed research outcomes, replicate the task classification independently, and distinguish assistance from additional output. It would also test whether the quieter areas—question selection, data gathering and experimental work—can become more productive without weakening scientific review.

For now, the report argues against the idea that AI has made research an end-to-end automated process. It shows a concentration around tasks that are already digital and comparatively easy to inspect. That may make today’s productivity gains real for some researchers, while leaving the hardest and most consequential parts of discovery stubbornly human.