Field diffusion · General science · Economics
Google and MIT map scientific AI use through Gemini telemetry, model citations and researcher surveys
Task-level measurements of scientific AI adoption, complementary model uses and reported workflow bottlenecks.
Summary
AI in Science: Early Insights combines roughly 15 million Gemini interactions, filtered to 360,000 scientific interactions, with an inventory of over 2,600 specialized models and a survey of 637 US/UK researchers. It finds complementary task profiles and reported bottlenecks in experiments and verification. Nearly seven weekly hours saved is a survey estimate.
AI role
Gemini interaction logs measure use; LLM classifiers assign scientific tasks. Specialized models supply predictions, while conversational models support analysis, coding and writing.
Narrative role
Adds direct usage telemetry to publication-based diffusion evidence and shows why faster computational tasks may leave experimental validation constrained.
Caveat
Single-vendor telemetry and inferred occupations/tasks limit representativeness. The survey is non-representative and self-reported; citations proxy use. Associations do not establish causal productivity gains or more discoveries.