AI in Science: Early Insights
A · Strong evidence Google, Google DeepMind and MIT FutureTech paper (co-authors include Daniel Rock, Neil Thompson and James Manyika) combining 15 million Gemini interactions, a 2,600-model inventory and a 637-scientist survey; strong data and method, held below S because the productivity figures are self-reported and the lead institution sells the tools.
Summary based on the full text·Original: about 70 minutes to read
In short
The first paper (September 2026) of a Google, Google DeepMind and MIT FutureTech programme tracking AI in science. Three sources are mapped to a new taxonomy of scientific tasks: about 15 million anonymised Gemini interactions (360,000 classified as science), an inventory of over 2,600 specialised scientific AI models since 2012, and a July-August 2026 survey of 637 researchers in the US and UK. Scientists use AI more than most occupations; nearly half use it daily. General LLMs and specialised models such as AlphaFold do different jobs with almost no overlap at fine task level. Scientists report saving almost 7 hours a week, but the bottleneck is moving downstream to lab work and validation, verification eats part of the gain, and many feel pushed toward safer questions.
Key numbers
- almost 7 hoursaverage weekly time scientists report saving with AI, mostly reinvested in research
- 41%surveyed scientists reporting a growing backlog of untested hypotheses
- 46%of those saving time who spend more than a quarter of it checking AI outputs (89% spend more than a tenth)
- 49% vs 28%scientists saying AI pushes them toward safer projects versus riskier ones
- about 2.7 timeslikelihood of US life, physical and social science occupations using AI relative to their employment share
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Life, physical and social science occupations are about 2.7 times more likely to use AI than their share of US employment predicts.
- Nearly half of papers introducing specialised models are in the top 1% of citations in their field; the team traced 460,000 citations to these models.
- A 1-point higher share of the world's researchers is associated with a 0.9-point higher share of science LLM interactions; some lower-income countries cite open models heavily without building them.
- LLMs are used for coding, statistics, literature review and drafting; specialised models for prediction, molecular design and simulation.
- Survey: about three quarters report time savings averaging almost 7 hours a week; about 8 in 10 report higher lab output over three years; 89% expect further increases; about 68% report more access to insights from other disciplines.
- More than 4 in 10 say their main constraint moved into lab execution, clinical validation or field data collection; 41% report a growing backlog of untested hypotheses.
- 89% of those saving time spend more than a tenth of it verifying AI outputs, 46% more than a quarter.
- 49% say AI pushes them toward safer, incremental projects; 28% say it lets them take on riskier questions.
How they got it
Gemini logs sampled in early April 2026 (excluding paid API use), filtered by a work classifier, occupation predictions and a new science classifier, then mapped to tasks with Google's OCTO clustering and classification engine. The model inventory combines agentic web search with Epoch AI's database and OpenAlex metadata. The survey covered principal investigators, industry researchers and early-career scientists across four domains.
Limits
- Log data covers Gemini only and is classified by Google's own models.
- Time savings, output gains and bottlenecks are self-reported.
- The survey covers the US and UK and 637 respondents.
- Authors are mostly employed by the company whose product is measured.
Why it matters for companies
The pattern applies well beyond labs. When AI speeds up the analysis and drafting steps, work piles up at the next stage: testing, approval, verification. R&D, legal and analytics functions should plan capacity for review and validation before expecting output gains to show up. The finding that verification takes a quarter or more of saved time for nearly half of users is a realistic discount to apply to any time-saving business case. The tilt toward safer projects is a management risk worth watching in innovation teams.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- AI enablement, AI adoption, Science