Scientists are using artificial intelligence to move faster through literature review, analysis, coding and hypothesis generation, according to a study published in September 2026. Yet the research suggests that these gains have created pressure later in the scientific process, where experiments, clinical validation and senior researchers' time cannot expand as quickly.
The study, a collaboration between Google ATLAS, Google DeepMind and MIT FutureTech, draws on approximately 15 million anonymized Gemini interactions, an inventory of 2,600 specialized scientific AI models and a survey of 637 active scientists in the United States and United Kingdom. It maps scientific work using a taxonomy developed by MIT FutureTech, covering tasks from reviewing literature to carrying out physical experiments.
Nearly half of surveyed scientists said they used some form of AI every day. Around three-quarters reported that AI had saved them time. Among researchers who reported a net saving, the average was just under seven hours per week.
The tools serve different parts of research. Large language models, or LLMs, are used for general analysis, coding and manuscript preparation. Specialized models are used for domain-specific prediction, data generation and classification. These specialized tools are relatively more common in health and life sciences, the study found.
Much of the time saved does not become available for entirely new work. Among scientists who said AI saved them time, about 89% said more than 10% of that time was spent verifying, debugging or fact-checking AI output. About 46% said those checks took more than one-quarter of the time they had saved.
The authors call this burden a 'verification tax', reflecting science's need for results that are reliable and correct. The study found that researchers primarily put time saved by AI back into further research, including developing and assessing additional ideas.
That reinvestment appears to be shifting the point at which research slows. As AI accelerates literature summaries, data analysis and the generation of possible research candidates, scientists can identify more questions that merit testing. The resources needed to answer those questions, however, have not grown in parallel.
About 44% of respondents said their main research bottleneck had moved downstream during the previous two years, towards work such as physical wet-lab activity or manuscript preparation. Physical experimentation, ethical approvals, clinical samples and the judgment of senior researchers were among the constraints identified in the study.
About 41% of respondents said their backlog of untested hypotheses had increased. Roughly 25% said it had fallen. The results point to a widening gap between the rate at which researchers can generate possible lines of inquiry and the rate at which they can validate them through experiments and other later-stage work.
Scientists use AI more than most other occupations, according to the research. Its findings nonetheless indicate that greater efficiency in conceptual and computational tasks has yet to produce a direct rise in the pace of breakthrough discoveries.
The study does not establish which disciplines face the greatest pressure, or whether new research structures will expand verification and experimental capacity. It also leaves open whether the verification burden will fall as scientists gain more experience assessing AI output. For now, its central finding is that AI has compressed the earlier stages of scientific work while physical and human limits remain in place later in the process.
Comments (0)