MIT and Google Release AI in Science Report: Hypothesis Generation Speeds Up, Experimentation Becomes New Bottleneck

The joint research team from Google, Google DeepMind, and MIT FutureTech released a 40-page report, "AI in Science: Early Insights," attempting to systematically assess the changes AI brings once it enters real scientific research settings. The report uses three data sets of rare scale: roughly 15 million anonymous Gemini interaction records, 2,690 specialized scientific AI models spanning different disciplines, and a survey of 637 active scientists in the United States and the United Kingdom, covering physics, computer science, life sciences, medicine, and social sciences.

Judging from the results, the changes AI brings are clearly two-sided. Nearly three-quarters of surveyed researchers say AI does save them time, on average close to 7 hours per week (elsewhere the report gives about 6.9 hours). But at the same time, more than 40% of scientists find the research bottleneck moving downstream in the workflow; 41% of respondents say they have more unverified hypotheses piling up than before.

This means AI can generate candidate answers faster and faster, but experiments, verification, and real-world data collection do not speed up in tandem. As the front end keeps accelerating, more and more hypotheses begin to pile up at the links that are hard to scale.

AI penetration in research is already quite high. Among the 637 surveyed scientists, about 46.6% use some form of AI every day, and another roughly 30.6% use it weekly. The report found that six categories of research-intensive occupations account for roughly 1.8 times the Gemini usage corresponding to their share of employment; looking only at core research positions such as life sciences, physical sciences, and social sciences, this ratio rises further to about 2.7 times. From roughly 15 million anonymous Gemini interactions, the research team also identified about 360,000 conversations that may be directly related to real research tasks.

The way AI enters science has long moved beyond the stage of "having a chatbot polish a paper." Among Gemini's research interactions, the highest share of tasks is quantitative data analysis and modeling, at about 42%, followed by research communication and writing; researchers also use it to handle research code, statistical analysis, literature reviews, experimental workflows, and even project management.

Another, more specialized scientific AI system is also growing rapidly. The 2,690 scientific models compiled in the report cover all 26 major scientific categories in the open academic database OpenAlex and 83% of subdisciplines, from AlphaFold, which predicts protein structures, to genomics analysis, materials discovery, medical imaging diagnosis, and weather prediction models, undertaking high-precision prediction, classification, simulation, and generation tasks. General-purpose large models and specialized models have not ended up replacing each other: at a coarse task-label level the two seem to both be doing "data analysis," but once the research workflow is broken apart, functional overlap quickly shrinks. They are more like a complementary division of labor—large language models are gradually becoming a general-purpose operating interface that threads research workflows together, while various specialized models go deep inside disciplines to handle vertical tasks.