Live·Open questions in longevity research
All news
Scientific ComputingScience Research

Anthropic analyzed 207 thousand research sessions with its AI assistant Claude: researchers use it for data analysis and writing, less often for ideas and data collection, which they consider most important; its estimates suggest that fully automating nonphysical work would speed up computer science by a factor of 15, while physical experiments would limit gains in biology and chemistry to a factor of 3–4

11 November 2026· 261111009

Anthropic analyzed 207 thousand research sessions with its AI assistant Claude: researchers use it for data analysis and writing, less often for ideas and data collection, which they consider most important; its estimates suggest that fully automating nonphysical work would speed up computer science by a factor of 15, while physical experiments would limit gains in biology and chemistry to a factor of 3–4

On November 5, Anthropic published a report on research work with Claude. A language model classified each research session from August and September by scientific field and task type. The sessions came from individual users of Claude and Claude Code, an agent that writes and runs code in users’ files. The report also includes a survey of 528 scientists and an economic model of how AI could accelerate research.

The report estimates that researchers spend 27,6% of their working time on data analysis, which accounts for 33,7% of Claude sessions. Writing accounts for 16,0% of working time and 32,0% of sessions. Data collection takes 28,6% of working time but accounts for just 11,5% of sessions, while ideas and hypotheses account for 7,1% of working time and 4,2% of sessions.

The survey respondents, mostly scientists in the life and health sciences who actively use AI, allocated 100 points across ideas, data collection, data analysis, and writing to rate their importance. Ideas received 35,6 points, data collection received 23,6, and data analysis received 23,3. Claude is used particularly rarely to collect clinical and biochemical data, where physical requirements limit its use. For ideas, the authors suggest that scientists may prefer to develop their own or that AI may be less capable of choosing research questions. In Claude Science, a version of Claude connected to scientific tools and databases, ideas account for 8% of sessions, twice the share in Claude.

The authors conclude that “scientists appear to use AI where it performs best.” They base this conclusion on a task ranking: a language model compared research steps in pairs, choosing the step in each pair that an AI agent working on a computer without laboratory access could perform better. Physical steps therefore rank lower in these comparisons. Tasks in the top 20% of the ranking, such as plotting data and running statistical tests, account for twice as many Claude sessions and four times as many Claude Code sessions as their share of working time would predict. Tasks in the bottom 20%, such as purifying reaction mixtures, see almost no use. Tasks rank higher when success is easy to verify and feedback is quick. These tasks are suited to reinforcement learning, in which a model repeatedly attempts a task and receives a reward for success.

In the authors’ model, research steps complement one another, so the overall gain is constrained by the steps where AI helps least. Substantial assistance with some steps only partly compensates for limited assistance with others. Using current AI wherever it is useful would increase research output by a factor of 1,35–2 for the same budget. In a hypothetical scenario in which AI performs all nonphysical work at roughly 1/16 of the cost of a researcher’s labor, output would increase by a factor of 15–16 in statistics and computer science, but only 3–4 in biology and chemistry. Physical steps account for about 40% of working time in those fields and become the largest cost once the remaining work is automated.

One of the lead authors, economist Lukas Althoff, wrote on X:

Physical bottlenecks may mean that scientists generate ideas faster than they can test them.

He cites a Google DeepMind survey in support: about 41% of scientists reported a growing backlog of hypotheses awaiting testing. He suggests that automated laboratories could potentially remove this constraint by speeding up physical testing of those hypotheses. Anthropic itself has a laboratory where its models can conduct physical experiments.

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#claude#research-automation#physical-bottlenecks#data-analysis#automated-laboratories