Anthropic report: Scientists use Claude most for data analysis and writing papers, and least for idea generation
On November 5, 2026, Anthropic released the research report AI for scientific discovery: How researchers use Claude, analyzing 207,049 research-related conversations under consumer plans from August 2 to September 30, 2026, including 128,914 conversations with Claude (Chat and Cowork) and 78,135 with Claude Code. Using privacy-preserving classifiers, the study retained only domain and task labels and did not store specific research content.
Data analysis accounted for 33.7% of Claude conversations and 48.1% of Claude Code conversations, compared with about 27.6% of scientists’ research time; writing and communication accounted for 32.0% of Claude conversations and about 16.0% of research time. Data collection accounted for about 11.5% to 11.9% of conversations and about 28.6% of research time; idea and hypothesis generation accounted for 4.2% of Claude conversations and only 2.1% of Claude Code conversations, versus about 7.1% of research time. Computer science accounted for only about 4.3% of researchers globally, yet made up 32.7% of research conversations in Claude Code.
The report also surveyed 528 scientists, mainly from the life and health sciences and active AI users, who ranked idea and hypothesis generation as the most important task, with data collection second. According to the report’s model estimates, if existing AI capabilities were widely adopted, research output in some fields could rise to about 2 times; if AI took over all non-hands-on steps, computer science and statistics could increase by about 15 times, while chemistry and biology, because about 40% of time requires hands-on experiments, could rise only about 3 to 4 times. In Claude Science, idea and hypothesis generation accounted for about 8% of conversations, and biomedicine and biology combined accounted for more than half of the conversations.
Original sources:
Title: AI for scientific discovery: How researchers use Claude Publisher: Anthropic (Johannes Hermle, Ben Jones, Lukas Althoff, et al.) Publication date: 2026-11-05 URL: https://www.anthropic.com/research/how-researchers-use-claude Source type: official report Verification status: This reading covered the webpage summary, key paragraphs of the main text, and the main chapters of the PDF full text; accessed on 2026-11-07 Supported content: conversation sample size and time period, task shares, comparisons with research time, survey results, relative use by field, estimates for acceleration scenarios, and Claude Science-related data
ABAB AI Insight
The report’s main author includes economist Ben Jones. Jones has long studied innovation and economic growth, and previously focused on the impact of technological progress on productivity at institutions including Northwestern University. In this study, he worked with Anthropic’s economics team to compare Claude’s actual usage data with the SciNet classification system, rather than relying only on self-reports. The report itself cites cases such as AlphaFold predicting more than 200 million protein structures and winning the 2024 Nobel Prize in Chemistry, and GNoME predicting about 380,000 stable new crystals, as background for AI’s involvement in scientific discovery, but these are breakthroughs in specific fields rather than the routine research workflow covered by this conversation statistics.
Claude usage was highly concentrated in data analysis and writing, in contrast to scientists’ self-rated “most important task” of idea and hypothesis generation. Low use for data collection mainly reflects the need for physical operations in clinical samples and biochemical experiments; low use for idea generation may reflect scientists’ preference to choose topics themselves, or the fact that AI is still weak at high-level research problem selection. The report notes that the top 20% of AI capability-rated tasks (such as data visualization and statistical testing) were used at about 2 to 4 times their share of research time, while the bottom 20% (mostly physical operations) were used almost not at all. This is related to task features that are easier to train with reinforcement learning and provide faster feedback, with a correlation coefficient of about 0.78. Scientists currently mainly use AI in areas where it is already strong and easy to verify.
Claude Science, a dedicated research workstation launched in late June 2026, connects to more than 60 scientific databases and tools. It is aimed at conversations in biomedicine and biology, which account for more than half of its usage; idea and hypothesis generation accounts for about twice the share seen in ordinary Claude, and the proportion of high-autonomy completions is also higher. This contrasts with general-purpose Claude: a dedicated tool can reduce friction in cross-database queries and bring more complex tasks into actual use. By contrast, experimental disciplines remain constrained by physical steps; even if all non-hands-on work were fully automated in the model, the productivity gains in chemistry and biology would still be far lower than in purely cognitive fields such as computer science and statistics.
This data shows that current AI mainly amplifies analysis and writing tasks where it is already strong, rather than directly replacing the topic selection and data collection that scientists value most. Because tasks in computing and statistics can be simulated extensively on computers and receive rapid feedback, their relative use and potential acceleration are higher; experimental fields, constrained by laboratory operations, have lower acceleration ceilings. The report presents model estimates rather than forecasts, and the sample is limited to individual subscription users and active AI adopters; actual output changes after broader adoption will still depend on the extent of dedicated-tool adoption and progress in experiment automation.
ABAB News · Cognitive Law
Scientists rate topic selection and data collection as the most important, while AI is used least there; analysis and writing far exceed their share of time.
The multiplier of AI use is positively correlated with task capability scores, and physical-operation steps have barely entered the conversations.
If non-hands-on work were fully automated, the acceleration ceiling in computer science and statistics would be about 15 times, while in experimental disciplines it would be only 3 to 4 times.