Anthropic study finds AI use in science tracks capability, with potential to double research output in some fields
Analysis of 207,000 research sessions shows scientists use Claude most for data analysis and writing, and less for idea generation or data gathering, aligning with AI's strengths.
1 source · cross-referenced
- Anthropic analyzed 207,000 Claude and Claude Code research sessions to study AI use across scientific fields and tasks.
- Scientists primarily use Claude for data analysis (one-third of sessions) and writing, with less use for idea generation or data gathering.
- AI use is higher in tasks and fields where AI capabilities are stronger, such as computer science and statistics.
- With current capabilities, research output in some fields could double; in a hypothetical scenario where AI automates all non-physical work, output in highly cognitive fields could rise up to 15-fold.
- About 40% of surveyed researchers expect AI models to carry out full research projects autonomously by 2030.
Anthropic published a research report analyzing 207,000 research sessions involving Claude and Claude Code to understand how scientists across fields use AI tools. The study uses privacy-preserving methods to categorize sessions by field and research task, relying on a predefined taxonomy to label activities such as data analysis, writing, idea generation, and data gathering.
The report finds that scientists use Claude most frequently for data analysis, which accounts for about one-third of sessions on Claude and nearly half of sessions on Claude Code. Writing is the second most common use case. In contrast, scientists use AI less for idea generation and data gathering, even though survey respondents ranked these tasks as critical for scientific discovery. The discrepancy suggests a gap between where AI is currently applied and where researchers believe it could have the greatest impact.
Usage patterns align with AI capability: tasks and fields where AI performs well—such as computer science, mathematics, and statistics—see higher adoption of Claude. The study introduces a measure of theoretical AI capability to explain these patterns, noting that AI use is two to four times higher in the top 20% of tasks it is best suited for compared to what would be expected based on researcher time allocation. Conversely, tasks where AI is least capable, including most physical work, see minimal use.
The report explores scenarios for how AI could accelerate research output. With current capabilities and widespread adoption, research output in some fields could double. In a hypothetical scenario where AI fully automates all non-physical research work, output in highly cognitive fields like computer science and statistics could increase as much as 15-fold. Fields constrained by physical bottlenecks, such as chemistry and biology, would see smaller gains, roughly three- to fourfold. The authors note that fields with the greatest potential for acceleration are also those that contribute to AI development, creating a potential feedback loop of faster AI progress.
Survey data from active AI users indicates that about 40% expect AI models to carry out full research projects autonomously by 2030. The report also acknowledges risks, including reduced human oversight, pursuit of unethical research, and shifts in researcher careers and motivation, emphasizing the need for careful consideration as AI becomes more integrated into scientific workflows.
- Nov 11, 2026 · The Verge — AI
Google DeepMind unveils AlphaGenome 2 to predict genetic variant effects on cell types
Trust79 - Nov 11, 2026 · Google DeepMind — Blog
Google DeepMind unveils AlphaGenome 2 to predict DNA variant effects in specific cell types
Trust84 - Nov 11, 2026 · arXiv cs.AI
Researchers demonstrate auditable human-agent teams for autonomous synchrotron X-ray experiments
Trust79