Anthropic report: AI use tracks RL-trainability, could speed science up to 15x

AGI Hunt · wechat · 2026-11-06

Anthropic's new report "How researchers use Claude" analyzes 207K research sessions (privacy-preserving) plus a survey of 528 scientists.

Key findings

Computer science is 4.3% of researchers worldwide but 32.7% of Claude Code research sessions; formal sciences use AI most, lab-based chemistry/biology and medicine least
Data analysis (33.7% of sessions) and writing (32%) dominate usage, while the tasks scientists rate most important — idea generation (4.2%) and data collection (11%) — are the least AI-assisted
Tasks in the top 20% of AI capability see 2-4x more usage than time-share predicts; capability scores correlate most strongly (0.78) with "RL fit" — tasks that are easy to verify get automated first

Acceleration estimates

With current capabilities broadly adopted, AI could take over 27-52% of research time (median 38%), yielding 1.35-2x speedups
Automating 75% of non-physical steps: 8x for CS; full automation: 15-16x for physical-work-free fields (statistics, CS, economics), but only 3-4x for biology and chemistry, where manual lab work is the bottleneck
The 16x ceiling stems from cost assumptions (AI $6/hr vs researcher $95/hr) — token prices literally cap research speed

The report also covers Claude Science, where protein-structure hypotheses appear 6x more often and 82.7% of sessions require PhD-level training.

More from AGI Musings