AI Can Take On 38% of Research Time, but Instrument Queues Won’t Get Shorter

Ethan Mollick shared a new Anthropic study. This Wharton professor has long tracked AI’s impact on work, and this time he set his sights on scientific research. The research team reviewed 207,049 research sessions from Claude and Claude Code, using a privacy-preserving method that had classifiers label only the field and task type without retaining the specific research content. The conclusion, in one sentence: at current capability levels, if widely adopted, AI can take on 27% to 52% of the researcher time needed to produce a new study, with a median of 38%.

This figure comes from the SciNet taxonomy, with 7,262 research tasks covering 34 fields and 318 subfields. The classifier tagged each session to a field and task, then compared that with the time researchers typically spend on various task types to calculate AI’s substitution share. The dataset includes 128,914 sessions from Claude and 78,135 sessions from Claude Code, spanning August 2 to September 30, 2026.

There is a clear imbalance in who uses it and where. Computer science accounts for 11.9% of Claude sessions and 32.7% of Claude Code sessions, while only 4.3% of researchers worldwide have a computer science background. Mathematics, statistics, and economics are also overrepresented. Lab science and medicine are underrepresented. Data analysis is the largest single use case, accounting for 33.7% of Claude sessions and 48.1% of Claude Code sessions. Writing accounts for 32% on Claude. Data collection and ideation are actually less common, at only 11.5% and 4.2%, respectively. Interestingly, in a companion survey, 528 researchers ranked ideation and hypothesis generation as the most critical tasks for important outcomes, with data collection second. The least-used tasks are exactly the ones they consider most important.

The research team created a theoretical AI capability score and rated each task. The result is clear: the top 20% of tasks by score received Claude usage 2 to 4 times their share of research time; the bottom 20% were barely used at all. The correlations between capability score and task characteristics are also straightforward: tasks suited to reinforcement learning, 0.78; cognitive tasks, 0.69; physical tasks, -0.57; interpersonal tasks, -0.31; tasks with regulatory constraints, -0.38. In short, AI is used more where it is good and less where it is not.

The acceleration scenarios fall into three tiers. With broad adoption at current capability, the arts are 1.35x, anthropology 1.36x, chemistry 1.41x, computer science 1.96x, and statistics 2.02x. If AI can automate 75% of non-physical steps, computer science reaches 8x, statistics 7x, economics 6x, most social sciences and mathematics 4 to 5x, and lab science 2 to 3x. If all non-physical steps are automated, statistics, computer science, economics, mathematics, psychology, and others could reach 15 to 16x, while biology, chemistry, materials, agriculture, and biomedical science only reach 3 to 4x. There is only one reason for the difference: physical steps. In lab science, 38% to 41% of research time is spent on physical steps, versus only 4% in the median field.

There is also a cost ceiling. The study assumes an AI agent costs $30 per hour and a researcher costs $95 per hour, and that for every hour of researcher time AI saves, it must run for 0.2 hours. That puts the cost of an automated task at roughly one-sixteenth of doing it oneself, so the speedup tops out around 16x. If AI costs halve, that ceiling doubles. The research team acknowledges that these are illustrative scenarios, not predictions; they do not account for queuing limits on large equipment, experimental material costs, or differences in the quality of AI-generated hypotheses.

Researchers’ expectations in the survey are not low. 40% think AI will be able to complete a full research project independently by 2030, 64% say by 2035, and 16% say never. The report also mentions a past episode: in a mid-2022 survey, experts were asked when AI would win a gold medal at the International Mathematical Olympiad; the median estimate was 2030, and superforecasters said 2035. In reality, it happened in 2025. The exception is tasks in the physical world, where the forecasting record is much better.

The comment section has a different tone. One commenter said the remaining 48% is for writing grant applications to pay for tokens. Another said that 38% of research time may be saved, but instrument queues won’t get shorter. Another said the fight over first authorship remains all too human. Someone added a jab: universities will find a way to turn that 38% into 38% more administrative work. These jokes all point to the same question: where does the saved time go? The report has no answer.

The report itself discusses several constraints. Adoption may be far slower than capability progress, and organizational restructuring takes time. Bottlenecks in physical experiments will not disappear just because AI gets stronger. A surge in paper volume could overwhelm peer review’s capacity to absorb it, or it could force improvements in research quality, depending on how journals and funders set incentives. There is also the alignment question: which questions are worth studying, how to study them, and whether humans remain in the key decision-making loop. These questions are far harder to answer than the 38% figure.

The value of this Anthropic study is not that it provides a definitive answer. It lays out how AI is actually used in research and puts the three variables—capability, adoption, and bottlenecks—in their proper places. 38% is a starting point, not an endpoint. Whether the saved time turns into more experiments, stricter validation, or more papers and more administrative forms depends on how the research system itself absorbs this wave of change.