AI for scientists is saving users nearly seven hours a week in self-reported time, but the extra speed is creating a less obvious problem: more ideas are reaching the parts of research that still move slowly. Instead of eliminating bottlenecks, AI appears to be shifting them downstream toward verification, physical experiments and clinical validation.
That is the most useful finding in a new collaboration involving Google, Google DeepMind and MIT FutureTech. The research combines a sample of 15 million Gemini interactions, an inventory of more than 2,600 specialised scientific AI models and a survey of more than 600 scientists in the United States and United Kingdom.
AI for scientists is already ordinary enough to measure
Nearly half of the scientists surveyed said they use some form of AI every day. That is notable because “AI in science” can sound like a specialist field involving a few advanced laboratories. The data instead suggest that general-purpose language models are being woven into ordinary research activities alongside specialist systems built for prediction, classification, simulation and other domain-specific jobs.
The research paper separates those roles. Large language models, proxied partly through Gemini usage, appear across general analysis, coding and manuscript preparation. Specialised models are more concentrated in scientific tasks where domain-specific prediction or generation is required. The researchers argue that the two forms of AI are often complements rather than substitutes.
That distinction is important. A scientist might use a language model to write code, restructure notes or explore an analysis, while a specialised model predicts a protein structure, classifies an image or simulates a physical process. Measuring “AI use” as one category hides those different jobs.
Seven saved hours do not automatically become seven hours of discovery
Surveyed scientists reported saving just under seven hours a week through AI, and the researchers say much of that time is reinvested in further research. The figure is self-reported, so it should not be treated as a clocked productivity measurement. People can overestimate or underestimate time savings, and the sample covers US and UK scientists rather than the entire global research community.
Even with that caveat, the pattern is revealing. When idea generation, coding or analysis gets faster, researchers can produce more candidate results and more hypotheses. But those outputs still need to be checked. A model can suggest an experiment in seconds; a laboratory may need days, weeks or months to run it. A candidate drug can be generated computationally far faster than it can be tested safely in a clinical process.
The result is a queue. The researchers report an increased backlog of untested hypotheses and substantial demand for output verification. Productivity therefore becomes a system property. Speeding up one stage of research can make the next slow stage more visible rather than making the entire scientific process equally fast.
LiveAIWire has already covered AI-assisted protein design, a field where computational systems can generate or rank possibilities at a pace that physical laboratories cannot simply mirror. The new study helps explain why rapid digital progress can coexist with stubbornly slow real-world validation.
The bottleneck moves to the expensive parts of science
Some scientific tasks scale cheaply because they happen in software. Running code on another dataset or asking a model to compare another set of papers can be repeated quickly. Wet-lab experiments, telescope time, specialised equipment, field studies and clinical trials are constrained by physical resources, people, safety procedures and time.
That means AI can increase demand for exactly the parts of science that are hardest to expand. A laboratory that can generate twice as many promising ideas may need more technicians, instruments and validation capacity, not fewer. Without that investment, researchers may end up with a larger pile of plausible work waiting to be tested.
This is not evidence that AI has failed to improve scientific productivity. It is almost the opposite. A new bottleneck can be evidence that an earlier constraint has loosened. The danger is assuming that a faster upstream step guarantees a proportionate increase in completed discoveries.
A similar issue appears in mathematical and theoretical work. LiveAIWire examined AI agents working on Navier-Stokes problems, where generating a possible line of reasoning is only part of the job. Verification, reproducibility and expert scrutiny remain central when the claim itself is difficult.
General models and specialist models are dividing the work
The inventory of more than 2,600 specialised AI models adds another useful piece to the picture. Public discussion often treats the latest general-purpose chatbot as if it were replacing every narrow scientific tool. The study instead finds broad use of both.
Specialised systems can encode data formats, scientific structures or training objectives that a general language model does not naturally possess. General models, meanwhile, can connect tasks around that specialist core: reading documentation, writing analysis code, translating results into prose or helping a researcher move between disciplines.
This combination may matter more than the performance of either category alone. A general model can lower the friction of using specialised tools, while the specialist model can provide capabilities that are difficult to reproduce through text prompting. Future research environments may therefore look less like one all-knowing AI and more like a network of models with different roles.
That resembles what is happening in education and professional training. LiveAIWire has reported on AI used as an instructor, where the tool can accelerate access to explanation without removing the need to test whether a learner has actually understood. In science, the equivalent question is whether faster assistance becomes validated knowledge.
The study measures use, not a new scientific golden age
There are important limits to the evidence. The researchers used Gemini interactions as one window into large-language-model use, which cannot represent every model or every institution. The survey records what scientists say about their own behaviour and time savings. The specialised-model inventory captures availability and use signals, not a direct causal measure of new discoveries.
The paper is therefore strongest as an early map of where AI is entering the workflow. It shows high adoption, different roles for general and specialised models, substantial perceived time savings and signs that pressure is moving to later stages of research. It does not establish that science as a whole is now seven hours per scientist per week more productive.
That distinction will become increasingly important as organisations try to measure returns on scientific AI. Counting prompts or model usage says little about whether experiments were completed, findings reproduced, papers improved or useful discoveries reached the outside world.
Faster science needs a redesigned pipeline
The most forward-looking implication is organisational. If AI makes hypothesis generation and analysis much faster, laboratories may need to redesign what comes next. That could mean automated experimental systems, better shared validation infrastructure, stronger data engineering or simply more human time allocated to checking outputs.
It could also change which skills are scarce. The person who can produce another candidate idea may become less valuable than the person who can design the decisive experiment, recognise an artefact or build a reliable validation process. AI can change the bottleneck and therefore change the economics of expertise.
The seven-hour figure is an attractive headline, but the backlog may be the deeper story. Productivity tools are often judged by how much faster they make an individual task. Science only advances when the whole chain works, from idea to evidence to verification. AI is already accelerating parts of that chain. The next challenge is making sure the rest can keep up.
Universities and research funders may also need to rethink what counts as capacity. If researchers can generate analyses and candidate hypotheses faster, the scarce resource may become access to instruments, technicians, curated datasets or replication teams. Funding another software licence could have less impact than expanding the laboratory stage that now receives the extra flow of AI-assisted ideas.
The same logic applies to publication. More drafts and analyses can increase the burden on peer review unless journals and research groups improve how claims are checked before submission. AI can therefore raise output upstream while creating a quality-control challenge downstream. The productivity story is not complete until the additional work survives those filters.
The researchers call their findings early insights, which is the right level of confidence. The important signal is not that AI has solved scientific productivity, but that enough researchers are now using it for second-order effects to appear. The location of the bottleneck is becoming measurable.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.