Scientists report saving nearly 7 hours a week with AI, and the bottleneck moves to verification

Mihai Codreanu, Alex Imas and coauthors draw on three sources: a sample of 15 million Gemini interactions, an inventory of over 2,600 specialized AI models across disciplines, and a survey of over 600 scientists. They map the data to a new taxonomy of scientific tasks.

Adoption is broad: scientists use AI more than most other occupations, and nearly half of those surveyed use some form of AI every day. LLMs and specialized models act as complements, with LLMs used for general analysis, coding and manuscript preparation and specialized models for domain-specific prediction, data generation and classification. Scientists report saving nearly 7 hours per week, time they mostly put back into research. As some stages get easier, bottlenecks move downstream: respondents describe a growing backlog of untested hypotheses and substantial demand for output verification.

The productivity figure is self-reported, and LLM use is proxied by Gemini usage alone. The study covers science broadly and does not break out finance or economics.