Artificial intelligence could multiply scientific output by up to 16 times in areas such as computing, mathematics and statistics, if it manages to automate all the intellectual stages of a research project. The projection is part of a study published by Anthropic on Thursday (5), which also identifies an important limitation: in the experimental sciences, physical work will continue to constrain productivity gains.
The company analyzed 207,049 research sessions in Claude and Claude Code, carried out between August 2 and September 30, 2026, and interviewed 528 scientists to understand how AI already takes part in academic work and which activities show the greatest potential for automation.
The survey revealed that researchers use the tools mainly to analyze data and produce texts. Data analysis accounts for 33.7% of scientific sessions in Claude and 48.1% in Claude Code. Idea and hypothesis generation appears in only 4.2% and 2.1% of interactions, respectively, despite being considered by the interviewed scientists the most important activity for new discoveries.
This difference helps explain why the automation potential varies so much between disciplines. In predominantly computational areas, a large part of the tasks can be carried out digitally. In biology and chemistry, even an AI capable of planning experiments and interpreting results still depends on the physical execution of these procedures.
Gains of 16 times are still hypothetical
Anthropic's most ambitious scenario considers that AI manages to fully carry out all the non-physical stages of a research project. Under these conditions, the economic model developed by the company estimates an increase of 15 to 16 times in scientific output in certain disciplines, keeping the research budget constant.
In the sciences that depend heavily on laboratories, such as biology, chemistry and materials science, the estimated gain falls to three to four times. According to the study, approximately 40% of the time of researchers in these areas is dedicated to physical activities, which would continue to require human work in the absence of automated laboratories.
With current AI capabilities, however, the estimates are considerably lower. Even assuming widespread adoption of the technology, Anthropic calculates that scientific output could grow 1.96 times in computing, 2.02 times in statistics and 1.41 times in chemistry.
These numbers represent modeling results, not productivity increases already observed. The simulations depend on assumptions about costs, distribution of tasks and automation capacity, without fully considering expenses with equipment and materials or the quality of the scientific hypotheses produced.
The report also identified signs of greater autonomy in specialized tools. In biology and biomedicine research carried out with Claude Science, 73.3% of sessions showed a high degree of AI autonomy, against 23.7% in conventional Claude. In addition, Anthropic estimated that 82.7% of tasks in Claude Science would require knowledge equivalent to a doctoral degree.
This autonomy, however, refers to the execution of specific tasks and does not demonstrate that the tool can already conduct complete scientific projects without human supervision.
Among the 528 researchers interviewed, predominantly from the life and health sciences and already familiar with AI, about 40% believe that models will be able to carry out entire research projects independently by 2030. The proportion reaches approximately 64% when the horizon is extended to 2035, while 16% consider that this capability will never be achieved.
