Anthropic study finds scientists use Claude mostly to analyse data and write

Anthropic has analysed 207,049 Claude and Claude Code sessions involving scientific work, finding that researchers use the tools most for data analysis and writing, and far less for generating ideas or gathering data. Its report also models how AI could increase research output, with the biggest projected gains in fields such as computer science and statistics.

Anthropic Watch analysis

What happened

The study covers consumer-plan sessions from 2 August to 30 September 2026: 128,914 from Claude and 78,135 from Claude Code. Anthropic says its researchers used classifiers to assign sessions to scientific fields and tasks, retaining those labels rather than the conversation details. Results were aggregated, with minimum account thresholds.

Computer science, mathematics and statistics show relatively high Claude use compared with their share of researchers. Data analysis accounts for about a third of Claude sessions and nearly half of Claude Code sessions. The report says scientists use the tools less for idea generation and data gathering, even though a separate survey found researchers rank those tasks among the most important for scientific discovery.

Key findings

• Use follows capability
  Tasks in the top fifth of Anthropic’s AI-capability measure get two to four times the Claude use that researchers’ time spent on them would predict.
• Current capabilities could make a difference
  Anthropic’s model suggests widely adopted current capabilities could double research output in some fields.
• The largest gains are hypothetical
  If AI could automate all non-physical research work, the report models output rising as much as 15-fold in computer science and statistics, versus three- or fourfold in laboratory sciences such as chemistry and biology.
• Researchers expect more autonomy
  Around 40% of AI-using researchers surveyed expect models to carry out full research projects on their own by 2030.

Why it matters

The report offers a useful snapshot of where AI is being used in research today, and where its use still falls short of researchers’ stated priorities. The gap between routine analysis and more ambitious idea generation is a reminder that adoption is not the same as scientific discovery.

The output figures are scenarios, not measured productivity gains or promises. The largest projections depend on AI taking over non-physical research work, while laboratory science still faces physical constraints. The study also measures Claude use on consumer plans, not AI adoption across all research or the quality of resulting work.

Our read

There is genuine signal here: researchers appear to use Claude more in fields and tasks where Anthropic’s measure says AI is better suited to the work. But a session count is not a breakthrough count. The most interesting finding may be the mismatch between what researchers do with these tools and what they say matters most for discovery. The productivity projections are worth debating, provided we keep the assumptions in the room.

What to watch

• Whether future research measures changes in the quality and pace of scientific results, not just tool use.
• Whether specialist tools shift more use towards idea generation and data gathering.
• How the model’s assumptions about automating research tasks hold up as AI capabilities change.

Discussion spark: Should scientific AI be judged mainly by how much research it helps produce, or by whether it advances the tasks researchers say matter most?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.