Google and MIT Release AI in Science Report: Scientists Save 7 Hours a Week

On September 15–16, Google, Google DeepMind and the MIT FutureTech team jointly released the first output of project Zvi, the "AI in Science: Early Insights" report (with an accompanying paper), using large-scale empirical data to characterize how scientists actually use AI in research. The findings are mixed: scientists save about 7 hours per week on average, but the nature of their work shifts toward validating results, and topic selection may drift toward "safer" questions.

Confirmed

• The report was produced by a cross-institutional team spanning Google, Google DeepMind and MIT FutureTech, with core members including Maria Codreanu (mcodreanu, lead), JMateosGarcia, alexolegimas and Daniel Rock.
• Data sources include 15M Gemini interactions, analysis of 2,600 specialized AI models, and a survey of roughly 600 scientists; JMateosGarcia's posts describe the report as combining multiple data sources.
• MITCSAIL and Ethan Mollick both relayed the core findings: scientists use LLMs and specialized AI models for different tasks, save about 7 hours per week on average, and reinvest most of that time in research—while new bottlenecks emerge, as saved time does not fully translate into additional output and validation burdens grow.
• Daniel Rock announced the paper on September 15, referencing a related thread by collaborator Arthur; a companion paper accompanies the report.
• This is the first deliverable of the team's long-term agenda on AI's impact on science.
• In follow-up discussion, an applied research organization noted the findings closely match its own experience: quantitative tasks are the most stable adoption area, and AI time savings shift workload to other parts of the workflow.

Why it matters

• This is among the first empirical works to systematically study scientists' AI usage based on real, large-scale Gemini logs, filling a data gap on how scientists actually use AI.
• The pattern of time savings offset by validation burden and more conservative topic choices aligns with practitioner experience, offering direct reference for research workflow design and policy.
• The Google–MIT cross-institutional collaboration with public data and methods provides a reusable research paradigm for future science policy and tooling.

2026-09-15 ~ 2026-09-16 · 18 related posts

• [source] Google's AI-in-science study mines 15M Gemini interactions, 2,600 models and 600-scientist survey — danielrock · 2026-09-15
• [source] Daniel Rock team releases new paper on how scientists use AI — danielrock · 2026-09-15
• MIT & Google study: AI saves scientists ~7 hours a week, new bottlenecks emerge — MIT_CSAIL · 2026-09-16
• [source] Google x MIT report: first insights on how scientists use AI, from Gemini logs — JMateosGarcia · 2026-09-16
• Behind the Google x MIT 'AI in Science' report: the full cross-org team — JMateosGarcia · 2026-09-16
• First large-scale 'AI in Science' report released as start of new research agenda — soumitrashukla9 · 2026-09-16
• Google DeepMind Releases AI in Science Report Based on 15M Gemini Interactions — JMateosGarcia · 2026-09-16
• Google and DeepMind release first AI in Science report on how scientists actually use AI — soumitrashukla9 · 2026-09-16
• Google Study: AI Saves Scientists 7 Hours a Week but Shifts Research Toward Safer Topics — emollick · 2026-09-16
• Study: Scientists Adopt AI Most in Quant Tasks, Saving Time but Shifting Work to Verification — JMateosGarcia · 2026-09-16

8 near-duplicate retellings: danielrock · soumitrashukla9 · MIT_CSAIL · JMateosGarcia · JMateosGarcia · JMateosGarcia · JMateosGarcia · JMateosGarcia