Scientific progress is the main driver of economic growth and prosperity. There is great excitement – but also concern – about the impact of AI on science, but so far there is little data. We provide preliminary insights into this from three data sources: a sample of 15 million Gemini interactions, an inventory of more than 2,600 specialized AI models across all disciplines, and a survey of more than 600 scientists. We map this data to a new taxonomy of scientific tasks to learn how scientists use AI. Four main findings emerged. First, we found broad adoption and coverage: scientists use AI more than any other occupation. Special AI models have broad disciplinary coverage and are highly regarded. Nearly half of the scientists surveyed report using some form of AI on a daily basis. Second, we document evidence that LLM (proxy through the use of Gemini) and specific models act as complements—LLM is used for general analysis, coding, and manuscript preparation, while specific models provide domain-specific predictions, data generation and classification. Thirdly, scientists report great productivity gains from using AI: saving almost 7 hours per week, time that is mainly reinvested in more research. Finally, we show that AI has changed the scientific process. As some stages of scientific research become easier, bottlenecks move downstream. Scientists report an increase in untested hypotheses and a growing demand to verify their output. Our findings show that AI has significant potential to increase scientific productivity. However, as in other sectors, the main impact will be regulated by the interdependence of complex tasks and investments to remove emerging bottlenecks.
It’s from a new paper by Mihai Codreanu, et.al.