A new study of AI in scientific work combines three unusually large but different windows: 15 million Gemini interactions, an inventory of 2,690 specialised scientific models and a survey of 637 active scientists in the United States and United Kingdom. The picture is not one universal “AI scientist”. General language models and domain-specific systems are doing different jobs.[1][2]
The study’s three evidence streams
| Measure | Value | Context |
|---|---|---|
| Gemini interactions | 15 million | Used to map how general-purpose AI appears in scientific tasks; Gemini is a proxy, not the whole market.Source 1 |
| Specialised scientific models | 2,690 | An inventory covering systems built for domain-specific prediction, simulation, classification and generation.Source 1Source 2 |
| Survey respondents | 637 scientists | Active researchers in the US and UK surveyed in July and August 2026.Source 3 |
| Reported weekly time saved | 6.9 hours on average | A self-reported productivity measure, not a timed experiment.Source 3 |
General models and specialised models appear to complement each other
Two tool families, different scientific jobs
| Tool family | Common role in the study | Example of the output |
|---|---|---|
| General-purpose language models | Analysis, coding support, literature work and manuscript preparation. | A script, explanation, draft or synthesis across text.Source 1Source 2 |
| Specialised scientific models | Domain prediction, data generation, classification and simulation. | A protein structure, materials candidate, image label or domain-specific forecast.Source 1Source 2 |
| Combined workflow | A language model helps organise the question and calls or interprets a specialist system. | A research loop in which general reasoning and specialised measurement remain distinct.Source 1 |
That division matters because it argues against one simple substitution story. A language model can help write code for an analysis without replacing the microscopy model that classifies an image. A specialised model can generate a candidate without deciding whether the experiment is worth running. Scientific work remains a chain of dependent tasks.
The saved time does not disappear—it moves the bottleneck
Survey respondents who used AI reported saving nearly seven hours each week and primarily reinvesting the gain in more research. If 6.9 hours were sustained for 52 weeks, it would equal about 359 hours, or almost 45 eight-hour working days. That extrapolation illustrates the scale of the claim; the study did not observe every respondent for a full year.
What acceleration seems to expose
| Reported pattern | Share or measure | Why it matters |
|---|---|---|
| Scientists using AI daily | 46.6% of respondents | Use is already routine for a large part of the sample.Source 3 |
| Primary bottleneck moved downstream | About 44% | Faster analysis can leave laboratory execution, clinical validation or field collection as the slower stage.Source 3 |
| Backlog of untested hypotheses grew | About 41% | Generating plausible ideas is becoming easier than checking them.Source 3 |
| Time-saving users who spend more than 10% of the gain checking outputs | 89% | Verification consumes part of the apparent productivity dividend.Source 3 |
| Time-saving users who spend more than 25% checking outputs | 46% | For many users, validation is not a marginal task.Source 3 |
Why this is an early map, not a causal verdict
The interaction sample comes from one provider. The specialised-model inventory depends on what the researchers could identify. Survey estimates depend on memory and perception, and frequent AI users may be more willing to participate. None of those sources proves that an hour saved produces a proportional increase in correct papers, successful experiments or useful inventions.[1][3]
The study’s most useful result is therefore the shape of the constraint. AI can accelerate literature review, coding and candidate generation while making physical experiments, expert review and replication relatively scarcer. A laboratory that buys models without expanding validation capacity may create a larger queue rather than a faster discovery system.
Sources & further reading
- AI in Science: Early Insights arXiv · 15 Sep 2026 · Accessed 16 Oct 2026
- AI in Science: Early Insights MIT FutureTech · 16 Sep 2026 · Accessed 16 Oct 2026
- Why AI is speeding up scientific research but not lab experiments Scientific American · 25 Sep 2026 · Accessed 16 Oct 2026