09-25-2026, 12:12 PM
“AI in Science: Early Insights” examines how artificial intelligence is already being used by scientists and what this means for the scientific research process. The authors combine three unusually broad sources of evidence: 15 million Gemini interactions, an inventory of more than 2,600 specialized AI models, and a survey of more than 600 scientists. The paper is an early empirical attempt to move beyond speculation about AI's effect on science and measure actual patterns of adoption and use.
The first major finding is that AI adoption among scientists is already substantial. Scientists use AI more than most other occupational groups, while specialized scientific AI models now cover a wide range of disciplines and have attracted significant citation attention. Almost half of the surveyed scientists report using AI every day. Importantly, AI is not confined to one particular type of scientific work: its applications range from general-purpose activities such as analysis, programming and manuscript preparation to highly specialized prediction, classification and data-generation tasks.
A particularly important result is that general-purpose LLMs and specialized scientific models appear to be complements rather than substitutes. LLMs are primarily useful for flexible, general tasks—coding, analysis, writing and information processing—while specialized models handle domain-specific problems such as prediction and classification. Scientists also report substantial productivity benefits: on average, AI saves them nearly seven hours per week, with much of that recovered time being redirected toward additional research rather than simply reducing working time.
The paper's most interesting conclusion is that AI may shift rather than eliminate bottlenecks in science. When AI makes certain stages of research faster—for example, generating hypotheses, analysing data or writing code—the constraint moves downstream. Scientists report an increasing backlog of hypotheses that have not yet been experimentally tested and greater demand for verification of AI-assisted outputs. Thus, the ultimate impact of AI on scientific productivity will depend not only on better models but also on complementary investments in experimentation, validation, infrastructure and human expertise. The authors describe the work as providing early insights, so the findings should be regarded as an initial empirical picture rather than a definitive long-term assessment.
Key takeaways
The first major finding is that AI adoption among scientists is already substantial. Scientists use AI more than most other occupational groups, while specialized scientific AI models now cover a wide range of disciplines and have attracted significant citation attention. Almost half of the surveyed scientists report using AI every day. Importantly, AI is not confined to one particular type of scientific work: its applications range from general-purpose activities such as analysis, programming and manuscript preparation to highly specialized prediction, classification and data-generation tasks.
A particularly important result is that general-purpose LLMs and specialized scientific models appear to be complements rather than substitutes. LLMs are primarily useful for flexible, general tasks—coding, analysis, writing and information processing—while specialized models handle domain-specific problems such as prediction and classification. Scientists also report substantial productivity benefits: on average, AI saves them nearly seven hours per week, with much of that recovered time being redirected toward additional research rather than simply reducing working time.
The paper's most interesting conclusion is that AI may shift rather than eliminate bottlenecks in science. When AI makes certain stages of research faster—for example, generating hypotheses, analysing data or writing code—the constraint moves downstream. Scientists report an increasing backlog of hypotheses that have not yet been experimentally tested and greater demand for verification of AI-assisted outputs. Thus, the ultimate impact of AI on scientific productivity will depend not only on better models but also on complementary investments in experimentation, validation, infrastructure and human expertise. The authors describe the work as providing early insights, so the findings should be regarded as an initial empirical picture rather than a definitive long-term assessment.
Key takeaways
- AI is already widely used in scientific research, rather than being merely an emerging possibility.
- Almost 50% of surveyed scientists use AI daily.
- The study analyses 15 million Gemini interactions plus data on 2,600+ specialized scientific AI models.
- LLMs and specialized AI systems complement each other rather than performing exactly the same functions.
- AI is being used for coding, analysis, writing, prediction, classification and data generation.
- Scientists report saving almost 7 hours per week through AI.
- Much of the saved time is being converted into more research activity.
- AI can make the generation of scientific ideas faster than the testing of those ideas.
- Verification and experimentation may become the new bottlenecks.
- The paper therefore suggests that AI's scientific impact depends on the whole research ecosystem, not simply on improving AI models.
- For mathematics, the broader implication is particularly relevant: AI may increasingly accelerate conjecture generation, computation, coding and proof exploration, while human mathematicians and formal systems remain important for verification and establishing correctness.
- This is an arXiv preprint submitted September 15, 2026, not a peer-reviewed journal publication at the time of submission.
┌────────────────────────────────┐
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘