Anthropic — How Scientists Use Claudeauto-ingested

Anthropic Economics report on 207,049 privacy-preserved research sessions on Claude, Claude Code and Claude Science. Scientists use AI where it works best: mostly for data analysis and writing, rarely for the ideation they rank most important. Scenarios put research acceleration at about 2× with today's AI and up to about 15× if all non-physical work is automated. Subscribed-feed capture (Anthropic Research), auto-fetched and auto-ingested 2026-11-08; awaiting human review.

Summary

Method. An LLM classifier labels each consumer-plan session (Aug 2 – Sep 30, 2026) against the SciNet taxonomy (Alamos & Althoff 2026): 7,262 tasks, 34 fields, 318 subfields. Only the labels are kept. There are 128,914 sessions on Claude (Chat + Cowork) and 78,135 on Claude Code, plus a supplemental Claude Science sample and a survey of 528 scientists (mostly life/health sciences, active AI adopters). Usage is benchmarked against researcher time per task (also SciNet, LLM-estimated and validated against wet-lab protocols and paper frequencies).

Thesis. "Scientists appear to use AI where it works best." Use tracks a measure of theoretical AI capability, which in turn tracks how cognitive a task is and how suited it is to reinforcement learning (AI Use Tracks AI Capability). The acceleration scenarios (AI-Accelerated Science) are capped by physical work, by AI cost, and possibly by ideation.

Key claims and data

Where use goes (fields).

• Field share of sessions tracks field share of researchers (log-log r = 0.79 on Claude, 0.62 on Claude Code), but formal sciences (CS, math, stats) and social sciences sit above the line; lab-based physical sciences and medicine sit below.
• Computer science is 4.3% of the world's researchers but 32.7% of Claude Code sessions. The AI/ML subfield is 1.2% of researchers, 5.0% of Claude and 15.2% of Claude Code sessions. Computational subfields lead in every domain.

Where use goes (tasks).

Task category	Claude	Claude Code	Researcher time
Data analysis	33.7%	48.1%	27.6%
Writing & communication	32.0%	—	16.0%
Data gathering	11.5%	11.9%	28.6%
Ideation & hypothesis generation	4.2%	2.1%	7.1%

• The survey matches the session data. Respondents rank ideation first and data gathering second for producing important results, yet both are underused. Data gathering is underused because much of it is physical. For ideation, the authors offer several readings: scientists may prefer to do it themselves, AI may be weaker at "choosing which research question to pursue," or the classifier may undercount it.

Claude Science (specialized tool). It connects Claude to fragmented research databases and tools. Biomedical sciences and biology are over 50% of its sessions. "Propose mechanistic models for protein–protein or protein–ligand interactions" is 5.6× more likely there than on Claude. Ideation is 8% of its sessions, about on par with researcher time. In bio/biomed, its tasks take about 2× the human hours, over 80% need doctoral-level training (against 50–70% elsewhere), and over 70% run with high AI autonomy (Claude 23.7%, Claude Code 64.1%).

Outputs. Written artifacts dominate in the social sciences and humanities, technical artifacts (code, data) in the natural and formal sciences. Technical artifacts are about 2× as common on Claude Code.

Use tracks capability. Capability scores come from pairwise "which step can AI do better?" judgments aggregated with a Bradley–Terry (Elo-style) model.

• Bottom 20% of tasks by capability (e.g. purifying reaction mixtures, participant observation) get almost no use. At the median, the use ratio is about 1. The top 20% (visualizing data, statistical tests) get about 2× expected use on Claude and 4× on Claude Code.
• Capability correlates with RL suitability (0.78) and cognitive work (0.69), and negatively with physical (−0.57), regulatory (−0.38) and interpersonal work (−0.31).
• Subfield use per researcher rises with average capability (0.45 Claude, 0.57 Claude Code). Health sciences use less than capability predicts (regulation, the authors suspect). Formal, social and humanities fields use more (text and code are easy to hand over).
• Higher-capability subfields also delegate more autonomously.

Acceleration scenarios (illustrative, "not forecasts"; ideas production function with complementary tasks, Jones 2026):

• Scenario 1, current capabilities widely adopted: AI takes on 27–52% of researcher time (median 38%). Output rises from 1.35× in the arts to 2.02× in statistics (CS 1.96×). Gains are limited because AI use is mostly assistive or shared, not autonomous.
• Scenario 2, 75% of non-physical steps automated: CS about 8×, statistics 7×, economics 6×; bench and medical sciences 2–3×; history, anthropology and the arts about 2× or less.
• Scenario 3, all non-physical steps automated: fields with little physical work reach 15–16×. Bench sciences stay at 3–4× because physical steps are about 40% of their researcher time.
• The cost ceiling: at $30/hr for AI vs $95/hr for researchers, with 0.2 AI hours per researcher hour saved, full automation costs about 1/16 of human work, so acceleration tops out near 16×. "Halving the cost of AI… would double the rate of progress."

Researcher expectations (survey). 38% say AI already matches top researchers at data analysis. By 2030, respondents expect AI to match the best researchers in data analysis (94%), writing (89%), and ideation and data gathering (72% each). About 40% expect AI to run full research projects by 2030 and 64% by 2035; 16% say never. The authors note that forecasters have underestimated AI progress (IMO gold arrived 5–10 years early), except in physical-world domains.

Discussion: six constraints.

1. Adoption lags capability: diffusion is slow, and gains wait for organizations to restructure (David 1990; Brynjolfsson et al.'s productivity J-curve).
2. Bottlenecks AI doesn't remove: physical experiments, and possibly choosing what to study.
3. AI could raise the quantity of research or its quality. A flood of papers could outstrip the community's ability to absorb and verify them.
4. Research organization could change, with AI doing refereeing and verification, or AI research communities running on their own.
5. Alignment: keep humans in key research decisions. Some risky domains, such as biology, may need deliberate checkpoints, accepting slower progress in exchange.
6. Researchers' jobs change. US PIs on federal grants spend about 44% of research time on administration, which is an attractive target.

Connections

• AI Use Tracks AI Capability — the report's central empirical finding, given its own page.
• AI-Accelerated Science — the scenarios, bottlenecks and constraints.
• The Bottleneck Shift — supporting evidence that ideation is underused and may be the binding constraint. It also challenges the claim that the bottleneck is durably human (see the contradiction there).
• Individual vs Organizational AI Gains — "adoption can substantially lag capabilities" independently backs Mollick's paradox.
• AI and Employment — Replacement vs Expansion — task-level evidence (AI takes over tasks, not jobs), plus the job-satisfaction question.
• Agent Autonomy Spectrum — the survey's none → assistive → shared → delegated → autonomous scale, and autonomy rising with capability and with specialized tools.
• Token Shortage — AI cost becomes the ceiling on acceleration once physical limits are gone.
• Human-in-the-Loop by Exception — deliberate checkpoints at key research decisions.
• Critical Thinking Is Domain-Bound — Claude Science's gains come from domain-specific tool integration, not from general capability alone.
• Method Reflection (T3) — the wiki already splits work the same way: AI does analysis and writing, and the human chooses the questions.
• Claude Code, Anthropic, Dario Amodei ("Machines of Loving Grace" is cited as the optimistic framing).
• Match Process Weight to Stage and Stakes seed — "deliberate checkpoints… accept slower progress to reduce risk" is a loose analogy, not grounding.

Source

• Original: raw/2026-11-05 Anthropic — How scientists use Claude.md · https://www.anthropic.com/research/how-researchers-use-claude · Appendix PDF linked from the article (not ingested).