Anthropic Research: How Researchers Use Claude for Scientific Discovery
Overview of AI in Scientific Research
AI is increasingly transforming the scientific process by amplifying human capabilities in cognitive research tasks. Anthropic’s recent study of 207,049 research sessions using Claude, Claude Code, and Claude Science reveals that scientists are already actively integrating AI into their workflows, with usage patterns strongly correlated to the theoretical capabilities of the models.
Usage Patterns Across Fields and Tasks
Claude usage is not uniform across the scientific landscape; it is heavily concentrated in formal sciences such as computer science, mathematics, and statistics. While general-purpose interfaces like Claude are used for a broad range of tasks, Claude Code is disproportionately utilized for computational work.
Across the research lifecycle, data analysis and writing are the most common AI-assisted tasks, significantly exceeding the time researchers typically allocate to these activities. Conversely, tasks involving data gathering and ideation—which researchers rank as highly important—remain underrepresented in AI usage. This gap is likely due to the physical nature of many data gathering tasks and the perceived limitations of current models in high-level hypothesis generation.
The Correlation Between Capability and Adoption
Anthropic developed a "theoretical AI capability" score for research tasks, finding that scientists use AI most where it is most capable. Tasks that are highly suitable for reinforcement learning (RL) and predominantly cognitive in nature see the highest adoption rates. In contrast, physical and interpersonal tasks, or those subject to strict regulatory barriers, see minimal AI integration. This suggests that as AI models improve their ability to handle complex, multi-step cognitive processes, adoption will likely expand into more specialized research areas.
Specialized Tools and Autonomy
Specialized interfaces like Claude Science demonstrate that domain-specific AI tools can bridge the gap between general AI and complex scientific workflows. Claude Science sessions are characterized by:
- Higher Complexity: Over 80% of usage requires doctoral-level training.
- Increased Autonomy: Over 70% of sessions are performed with a high degree of AI autonomy.
- Task Specialization: Users employ these tools for intricate tasks like proposing mechanistic models for protein-ligand interactions, which are rarely seen in general-purpose chat sessions.
Potential for Research Acceleration
Based on current capabilities and survey data, Anthropic estimates that widespread AI adoption could double research output in fields like computer science and statistics. In more speculative scenarios where AI automates all non-physical research steps, output in cognitive-heavy fields could theoretically increase up to 15-fold. However, physical bottlenecks remain a significant constraint; bench sciences like biology and chemistry, which rely on physical experimentation, are projected to see more modest gains (3-4x) unless accompanied by advances in automated laboratory infrastructure.
Future Outlook and Risks
Researchers are optimistic about the future of AI in science, with approximately 40% expecting AI to carry out full research projects autonomously by 2030. While these advancements promise to accelerate human welfare through faster discovery, they also introduce new challenges. These include the need for human oversight to ensure safety and ethical alignment, the potential for AI to overwhelm the scientific community with increased output, and the necessity of restructuring research organizations to fully realize the productivity gains promised by these new tools.