Google, Google DeepMind and MIT FutureTech published a study on AI’s impact on science. It analyzed 15M interactions, 2,690 models and surveyed 637 researchers. Researchers report saving about 7 hours a week using AI. Nearly half now face a growing backlog of untested hypotheses. 44% say their main bottleneck has moved downstream. 88% spend over 10% of saved time verifying AI outputs. Nearly half spend over a quarter of saved time on verification. Computational fields can close the entire loop digitally. Physical sciences face constraints in experimentation and data collection. AI may encourage safer, more incremental projects. Mihai Codreanu co-led the study and discussed its findings. He expects verification to remain a human task for now. AI could help prioritize hypotheses using scoring frameworks. The study explores how AI shapes research agendas and discovery processes.

• 15M Gemini interactions analyzed
• 2,690 specialized models surveyed
• 637 US and UK researchers surveyed
• 7 hours saved per week