Scientists use AI for the grunt work. The parts they think matter most, they still do themselves.
Jump to a section
See the story at a glance
There's a thing that happens when you ask researchers what matters most in science. They say: generating new ideas, formulating hypotheses, figuring out what to study next. The creative step. The judgment call. The part that's actually hard and consequential.
Then the Anthropic Institute watched 207,049 sessions of scientists actually using Claude. What those scientists were doing was mostly different. Data analysis: 33.7% of Claude sessions, versus 27.6% of their estimated time on that task. Ideation — the thing they said matters most? 4.2% of Claude sessions, versus 7.1% of their time.
Scientists use AI for the grunt work. The parts they say matter most for discovery, they still do themselves.
That gap is not a criticism. It might just be honesty about where AI is actually useful right now. And it's the most interesting thing in an otherwise technically dense November 5 report from lead authors Johannes Hermle, Ben Jones, and Lukas Althoff.
Source spread
- Anthropic Institute — "AI for scientific discovery: How researchers use Claude" (Nov 5, 2026) — academic. 207,049 consumer-plan sessions (Claude Chat, Cowork, and Code) from August 2 to September 30, 2026, labeled using a 7,262-task research taxonomy. First-party data, transparent methodology, honest about what the acceleration estimates are (illustrative assumptions, not forecasts).
- 528 surveyed researchers — builder. Life and health sciences primarily. Active AI adopters, surveyed about tasks, importance, and expectations. Sampling on AI enthusiasm is worth noting.
Pros & cons
What the data shows:
- AI usage is concentrated where AI is actually capable. The correlation between task AI-capability scores and usage rates is tight — the top 20% of tasks by AI capability see roughly two to four times expected usage. That's not random adoption. That's researchers having tried things and gravitating to what works.
- Computer science researchers are dramatically overrepresented: 4.3% of the world's researchers but 32.7% of Claude Code sessions. This is partly self-selection (CS researchers are more comfortable with the tools) and partly a signal about which fields have AI-capable workflows available.
- The acceleration scenarios are specific. With current capabilities, the model implies 27–52% of researcher time could go to AI (median 38%), implying 1.35× to 2.02× research output. If AI automated all non-physical steps, computer science and statistics could reach roughly 15×, while biology and chemistry would stay around 3–4×.
What deserves skepticism:
- The acceleration scenarios use stated assumptions: $30 per AI agent hour vs $95 per human researcher hour, 0.2 AI hours per human hour saved. The paper says explicitly these are illustrative, not forecasts. The 15× CS figure is arithmetic applied to those inputs — a sensitivity analysis, not a prediction.
- The 40% figure ("expect AI to run full research projects by 2030") comes from the surveyed sample of active AI adopters, not a representative cross-section of researchers. Sampling on enthusiasm produces enthusiastic projections. That doesn't make the number wrong. It does mean you should read it as an upper bound from optimists.
- Physical bottlenecks are explicitly excluded from the model. For biology and chemistry, those bottlenecks — the time it takes to actually run an experiment — are often the real constraint on research velocity. The paper flags this.
What builders need to know
- If you're building research tools, data analysis and writing are where researchers are already comfortable with AI and where your tool will get the most organic adoption. Ideation and data gathering are the opportunity gap — the tasks where researchers say the value is highest but AI use is lowest. If you can close the trust or capability gap there, you're in less-crowded territory.
- Claude Science shows a different profile than general Claude: tasks are longer, more autonomous, and over 80% require doctoral-level knowledge, with biomedical sciences and biology dominant. That's the direction specialized scientific tools are moving — away from general chat, toward longer-horizon agentic work on hard technical problems.
- The acceleration model assumes $30 per AI agent hour. If you're pricing research tools, that's the implicit ceiling on the perceived value proposition. Teams that think AI hours cost more than that will resist automation even if the tools work well.
- The session-classification taxonomy used in the study covers 7,262 tasks across 34 scientific fields — described in the methodology section of the report. It's the most detailed public map of what scientific research work actually consists of. Worth reading if you're building in this space.
Further reading
- Anthropic Institute — "AI for scientific discovery: How researchers use Claude" (Nov 5, 2026) — full report with methodology, task taxonomy, and acceleration model
- Anthropic AI for Science program — expansion of the program supporting researchers with Claude credits, expanded to 20 foundations and institutions as of Oct 28, 2026
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Claude designed a speakerphone. Engineers built it. It still works.
Anthropic asked Claude Fable 5.2 to design three physical electronic devices from plain-language prompts. Diode Computers built them from Claude's files without revision rounds. After weeks of use, all three prototypes still work. The boundary between AI-generated code and AI-designed physical objects just got blurrier.
The piece of agent architecture most builders are skipping: the dreaming step.
Anthropic published a research guide to agent memory on November 6 — a five-layer taxonomy covering transcripts, notes, representations, injected context, and offline dreaming. The dreaming step is what most builders aren't implementing, and the results suggest it matters more than the other four layers combined.
Anthropic proved the model didn't lie. The proof cost 72× as much as the inference.
Anthropic's Frontier Red Team published FlashProve on November 2 — an open-source zkSNARK system that lets a verifier confirm a specific committed model produced a given response. The 72× compute overhead and 1,024-token context cap put production use years away, but the direction is correct.