A new study puts hard numbers on something everyone suspected: scientists have quietly become some of the heaviest AI users in any profession.
Researchers analyzed three data sources - 15 million interactions with Google's Gemini, an inventory of more than 2,600 specialized AI models built for specific disciplines, and a survey of over 600 scientists - to map how researchers actually use AI day to day. Nearly half of surveyed scientists say they use some form of AI every day, more than workers in most other occupations report. The data reveal a clear division of labor: general-purpose LLMs like Gemini handle coding, literature analysis, and manuscript drafting, while specialized models handle domain-specific prediction, data generation, and classification. Scientists reported saving close to 7 hours a week, time they said they mostly reinvest in more research rather than bank as free time.
The productivity gains are real, but they are not evenly distributed across the research pipeline. As AI speeds up early-stage work like literature review and hypothesis generation, the bottleneck shifts downstream: scientists report a growing backlog of untested hypotheses and rising demand for someone, or something, to verify AI-assisted output.
The takeaway isn't that AI is transforming discovery so much as it's rerouting the bottleneck: hypothesis generation speeds up while verification becomes the new choke point.