Anthropic study finds scientists use Claude mostly for data analysis and writing
Anthropic reports that Claude use across scientific fields and tasks tracks what AI can theoretically do, with data analysis and writing dominating.
Anthropic's Economics team reports that scientists using Claude and Claude Code concentrate on data analysis and writing. Where they use the tools tracks how well AI can perform each research task in theory. In a sample of 207,049 sessions from August 2 to September 30, 2026, data analysis made up 33.7% of Claude sessions and 48.1% of Claude Code sessions. Writing and communication made up 32.0% of Claude sessions.
A language model labeled each session by field and research task using the SciNet taxonomy, and only those labels were kept for analysis. The report says no classifier stored open-ended details such as datasets, hypotheses or unpublished findings. Results are published only in aggregate, for groups with enough distinct accounts. To judge whether a share of use is high or low, the team compared it with the share of researcher time that SciNet estimates each task takes. By that benchmark, data analysis is overrepresented. It accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions, against 27.6% of researcher time. Writing and communication is 32.0% of Claude sessions against 16.0% of researcher time. Data gathering runs the other way: it takes 28.6% of researcher time but only 11.5% of Claude sessions and 11.9% of Claude Code sessions. Ideation and hypothesis generation takes 7.1% of researcher time and accounts for 4.2% of Claude sessions and 2.1% of Claude Code sessions.
The field pattern follows a similar logic. Compared with each field's share of the world's researchers, the formal sciences use Claude heavily, while health and lab-based physical sciences use it less. Computer science is 4.3% of the world's researchers but 32.7% of Claude Code sessions. Artificial intelligence and machine learning is 1.2% of researchers yet 5.0% of Claude sessions and 15.2% of Claude Code sessions. Medicine is the third-largest field by Claude session volume, but the report says its relative use is low, since researchers spend much of their time with patients. A footnote adds that sessions are assigned to a field by the research being pursued rather than by the tool used. As a result, coding done for other computational sciences is not counted as computer science.
The report's central argument is that usage follows capability. The mechanism is worth understanding. The team asked an LLM classifier to compare pairs of task steps and say which one AI could perform better. It then aggregated those head-to-head judgments with a Bradley-Terry model, the same family of model behind chess Elo ratings. It averaged step scores into a task score it calls theoretical AI capability. Next, it computed a use ratio: a task's share of sessions divided by its share of researcher time. A ratio of 1 means usage matches time spent. The top 20% of tasks by capability score draw about twice the expected use on Claude and about four times on Claude Code. The bottom 20%, which includes most physical work, are barely used. The median task sits near 1.
The team also correlated the capability score with other task labels. The score correlates at 0.78 with suitability for reinforcement learning, 0.69 with cognitive work, -0.57 with physical work, -0.31 with interpersonal work and -0.38 with regulatory restriction. The report calls the score theoretical. It also notes that health sciences use Claude less than their capability scores would predict. It suggests regulation of medical data as a possible reason, without testing that explanation.
The survey adds a separate check. The team asked 528 scientists, mostly in the life and health sciences, how important each task category is for an important result. Respondents ranked ideation first and data gathering second, even though their self-reported AI use for both is lower than for data analysis and writing. The participants were recruited from applicants to Anthropic's AI for Science and related grant programs, and are affiliated with institutions in the US and Canada. A footnote says that because they applied for AI research credits, they are likely more intensive AI users than typical scientists. The report offers two readings of the gap. Data gathering is largely physical, while ideation is not, so the gap there has a different cause. Ideation may also be undercounted in sessions, because it is woven into other tasks and hard to classify. The report also concedes that AI may be genuinely weaker at high-level ideation, such as choosing which question to pursue.
The report also looks at Claude Science, a specialized tool. Biomedical sciences and biology account for more than 50% of its sessions. Ideation and hypothesis generation make up 8% of Claude Science sessions, roughly on par with the share of researcher time these tasks take. The report says the task of proposing mechanistic models for protein interactions is 5.6 times as likely to appear on Claude Science as on Claude, because the tool can pull structures and binding data from connected databases. Sessions on Claude Science would take nearly twice as many human hours to complete without AI as work on Claude and Claude Code. Over 80% are estimated to require doctoral-level training. Footnote 8 cautions that these differences may partly reflect who adopts the tool, since adopters are likely to have more demanding use cases.
The final section models how much faster research could go. In a scenario based on current capabilities and survey-derived levels of AI use, AI could take on 27 to 52% of the researcher time needed for a study, with a median of 38% across fields. That translates into accelerations from 1.35x in the arts and 1.41x in chemistry to 1.96x in computer science and 2.02x in statistics. If AI automated 75% of non-physical steps, computer science would reach about 8x. If it automated all non-physical steps, fields with little physical work would reach about 15 to 16x. Biology, chemistry, materials science, agricultural sciences and biomedical sciences would reach only 3 to 4x, because lab work remains a bottleneck. The model assumes an hour of AI agent time costs $30 against $95 for an hour of a researcher's time, and that AI runs 0.2 hours for each researcher hour it saves. The report calls these illustrative and says the AI cost uses today's public list prices. Its ceiling of roughly 16x follows from those assumptions, since a fully automated task costs about one-sixteenth of researcher time.
The same survey asked researchers when they expected AI to match top researchers working alone. Over half expected that for data analysis by 2027, and 38% said it had already happened. By 2030, 94% expected it for data analysis, 89% for writing, and 72% each for ideation and data gathering. About 40% expected AI to carry out full research projects by 2030, about 64% by 2035, and 16% said it would never happen. These are expectations from the same adopter-heavy sample, not measured outcomes.
The report lists what its model leaves out. It does not model constraints at the level of a whole field, such as shared facilities, nor the cost of machinery and experimental inputs. It does not count simulations as a substitute for physical experiments, and it does not account for the quality of ideas. The authors write that if AI identifies better hypotheses, research could speed up further. If research taste stays uniquely human, progress would be slower. The acceleration figures therefore measure how much more output a fixed budget could produce, not how much better the resulting science would be.
What changed
Scientists' use of Claude and Claude Code concentrates on data analysis and writing, while data gathering and ideation are used far less than their share of researcher time.
Type: Research · Applies to: Claude, Claude Code
Why it matters
The report links AI use to task capability, suggesting the largest near-term gains are in cognitive, text- and code-based work rather than physical research. Organizations in lab-heavy fields should expect smaller gains than in computational ones.
What you can do
Map your own research or analytical workflows by task type to see which steps are cognitive and text- or code-based, where the report suggests AI is most useful.
Original source
Where we saw it
Anthropic Research
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works