Researchers Use Claude Where It Fits Best, Not Where They Value It Most

Scientists are using AI where it fits best, not where they say it matters most. Anthropic’s Economics team studied 207,000 privacy-preserved Claude and Claude Code research sessions and found data analysis makes up about a third of Claude sessions and nearly half of Claude Code sessions. Idea generation and data gathering get less use, even though surveyed scientists rank them as the most important tasks.

The pattern follows task fit. Tasks in the top 20% on a theoretical capability measure see two to four times the use that researcher time alone would predict. Use is heaviest in computer science, mathematics and statistics, and lighter in health and lab-based physical sciences.

That matters for any company deciding where to point AI. Adoption tracks what the tools handle well today, so the first wins land in analysis-shaped work while the judgment-heavy work lags.

The gains can be large when the fit is right. Anthropic’s own biology lab reports that Claude helped produce the first 3D structure of a vault particle from a prokaryote: a purification that conventionally takes three days took about 90 minutes, and a first structural model arrived 34 hours later, which Anthropic estimates would normally take weeks. Speed on the steps did not settle the open question, though — the function of vaults remains unresolved.

The report’s scenario model is measured. Widely adopted current capabilities could double research output in some fields. About 40% of surveyed AI-using researchers expect models to run full projects on their own by 2030, and the authors flag oversight risks.

The takeaway for operators — measure which tasks your people already hand to AI, compare that with the tasks they say matter most, and ship against the gap. Outcomes, not tool counts, are the test.