Scientists ask AI for hypotheses, get spreadsheets instead

Science got a fast assistant — and no co-author. Anthropic published a report on how researchers actually use Claude, and the numbers hit the main promise of AI in science: that the machine will do the thinking.

🔬 What they found

A gap between words and logs. In the survey, hypothesis generation ranks first among the activities that matter for discovery. In real sessions it takes about 4% of the work. The rest is data analysis, code, editing, drafting text.

People bring the model where something needs to be crunched and cleaned. They don't bring it where something needs to be invented.

⚙️ How it works

Anthropic didn't ask what researchers think about AI. It pulled logs: 207,049 sessions in Claude and Claude Code between August 2 and September 30, sorted them by task type, and compared them with answers from 52 scientists.

Then arithmetic: how many sessions per task, what people type, how that lines up with what they call important. It doesn't line up.

🧭 How it differs

Typical "AI in science" studies are benchmarks on tasks or counts of mentions in papers. This is a map of habits: what people actually type into the box at work.

And one honest caveat: a vendor is reporting on its own product. Not independent research. Data from inside.

📊 What it showed

The gap matters more than the percentage. Hypotheses lead the survey and trail the logs. Data analysis and text do the opposite.

Claude Code is a separate lane: an engineering tool to compute, assemble, run. Nobody orders ideas there.

🌍 What it means for AI

If hypotheses are the bottleneck, the next lab product won't be another chat window. It'll be a loop: literature, data, a draft hypothesis, automated checks, filtering.

Whoever builds that loop first takes science. For now, AI optimizes what is already measurable. Researchers feel that — and keep the most valuable part to themselves.

⚠️ What is not proven

One vendor, one product, two months. The survey is 52 people — a conversation, not a sample. No peer review.

The 4% is a share of sessions, not a verdict on hypotheses. Maybe researchers just don't know how to phrase a prompt that gets a decent idea. And nobody checked how many machine-proposed hypotheses later held up. Without that, the argument goes nowhere.

Prediction: Within a year, at least one major lab tool will sell a closed loop of hypothesis and verification rather than a chat box.

Disclaimer: vendor report on its own data, not peer-reviewed

Anthropic Blog · AI Times