Is AI Making Scientific Research More Productive or More Valuable?

Pretty much everyone knows that AI is helping scientists get more work done in less time. Researchers can analyse huge amounts of data, review existing studies and develop new ideas much faster than they could before. But just because scientists can work faster, does that mean they're making more important discoveries? 

New research from Anthropic suggests the answer isn't quite that simple. Its November 2026 study found that scientists are using AI heavily for tasks like data analysis and writing, but far less for activities such as coming up with new scientific ideas. And there's another question worth asking. 

If AI helps a researcher produce twice as much work, how do we know whether any of that extra work is actually better? For organisations investing in AI in scientific research, understanding the difference between getting more work done and producing better results is becoming an important part of deciding where to spend their money.

Where AI Is Improving Scientific Productivity

AI is already changing the way scientists work, particularly when they're dealing with large amounts of information or tasks that take a lot of time. Think about how much work goes into preparing for a scientific experiment. Researchers often need to read dozens of existing studies, organise data, write code and analyse earlier results before they can even begin testing a new idea. 

AI can now help with many of these tasks, leaving scientists with more time for the research itself. The Anthropic Institute's November 2026 study looked at 207,049 scientific research sessions using Claude and Claude Code. It found that data analysis accounted for 33.7% of research sessions on Claude and 48.1% on Claude Code. 

Writing and communication were also common uses. But only 4.2% of Claude sessions and 2.1% of Claude Code sessions involved coming up with new ideas or scientific hypotheses. A hypothesis is simply a possible explanation that scientists can investigate and test. 

Interestingly, when Anthropic surveyed 528 researchers, they identified coming up with hypotheses and gathering data as some of the most important activities for producing significant scientific results. So scientists are using AI most often to help with the work surrounding scientific discovery, rather than the activities they believe are most likely to lead to important findings. 

Google's September 2026 AI in Science research, developed with Google DeepMind and MIT FutureTech, found that scientists reported saving almost seven hours a week by using AI. That's a considerable amount of time they could spend on other research activities, although the savings were reported by the researchers themselves rather than measured independently. 

Anthropic also looked at how much faster scientific research could become if AI were used more widely. Its estimates suggest that current AI capabilities could potentially double research output in some fields, with even greater improvements possible as the technology develops. 

But these are predictions based on different scenarios, not proof that scientists are already making discoveries twice as quickly. And while getting more work done is certainly useful, there's still a difference between producing more research and producing research that actually moves science forward.

Why More Research Doesn't Always Mean Better Discoveries

Scientific progress isn't just about how many experiments researchers complete or how many papers they publish. Those things help scientists share knowledge and test ideas, but they don't tell us much about how useful or reliable the results are. AI can help researchers produce more work without necessarily improving what they discover. 

At the same time, there are cases where AI has helped scientists investigate problems they couldn't easily solve before. The difference becomes clearer when we look at what researchers are actually measuring.

Research output isn't the same as research quality

Imagine two scientists working on similar problems. One publishes ten research papers in a year, while the other publishes five. Looking at those numbers alone, you'd probably assume the first scientist was more productive. But what if the second scientist's work led to an important discovery, while most of the first scientist's papers added very little new knowledge? 

That's the problem with using research output as a measure of scientific progress. A January 2026 study published in Nature looked at 41.3 million scientific papers and found some interesting differences between scientists using AI and those who weren't. Researchers whose work involved AI published 3.02 times more papers and received 4.84 times more citations. 

Citations are references to a study in other research papers, so they're often used to measure how much attention scientific work receives. But the study also found that AI adoption was linked to a 4.63% reduction in the range of scientific topics being studied. 

In other words, individual scientists appeared to be producing more research and receiving more attention, while the wider scientific community was concentrating on a smaller range of subjects. The study doesn't prove AI caused those changes, but it does show why publishing more papers isn't necessarily the same as making broader scientific progress. 

Other research has found more encouraging results. A July 2026 study published in Research Policy examined more than 80 million scientific publications across over 170 fields. It found that research involving AI was generally linked to more original findings and greater scientific impact, although the results varied considerably between fields. 

Both studies tell us something useful. AI can help scientists produce more work and, in some cases, more original research. But neither publication numbers nor citations can tell us everything about the importance of a discovery.

Some productivity gains do lead to valuable discoveries

There are also examples of AI doing more than simply helping researchers work faster. Scientists have used AI to predict protein structures, identify promising new materials and investigate problems that would otherwise involve enormous amounts of manual work. Take protein structures, for example. 

Proteins are essential to how living things function, and their three-dimensional shapes can tell scientists a lot about what they do. Working out those shapes through traditional experiments can be difficult and time-consuming. 

AI tools such as AlphaFold can predict them, giving researchers a useful starting point for investigating how proteins work or interact with other molecules. Of course, predicting a structure isn't the same as proving that it's correct. Scientists still need to check the results and understand what they mean for the problem they're studying. 

And AI doesn't remove the need for scientific experience. Researchers still have to recognise promising ideas, question unexpected findings and decide which results are worth investigating further. So AI can help scientists make meaningful discoveries. 

But the results depend on what researchers are trying to achieve, how they use the technology and what happens after AI has done its part.

What Determines Whether AI Improves Scientific Outcomes?

Scientific research involves a lot more than coming up with an idea and testing it. Researchers need to gather information, develop possible explanations, design experiments, check their results and work out what they've actually discovered. All these activities are connected, which means speeding up one part of the process doesn't necessarily make the entire project move faster. 

For organisations investing in AI for research and development, the benefits often depend on where AI is being used and whether the rest of the research process can keep up.

Research bottlenecks don't disappear when tasks get faster

Imagine a team of scientists trying to understand how a particular disease works. They use AI to analyse existing research and come up with dozens of possible explanations. Previously, developing that many ideas might have taken weeks or even months. Now they have all those ideas in a fraction of the time. But there's a problem. 

Each one still needs to be checked, and some will require laboratory experiments, specialist equipment or information the researchers haven't collected yet. If the team can only test five ideas a month, having fifty new hypotheses doesn't suddenly mean they can make ten times as many discoveries. 

Google's September research found signs of exactly this kind of problem. Scientists reported growing numbers of untested hypotheses, along with more time being spent checking AI-generated work. It also helps explain why Anthropic's predictions for research acceleration differ between scientific fields. 

Computer science and statistics involve a lot of work that can be completed digitally. Chemistry and biology often depend on physical experiments that take time, equipment and resources, regardless of how quickly AI produces the initial analysis. So while AI can remove some of the work slowing scientists down, other tasks can become the new bottleneck.

The value of AI depends on the research being done

Not every research project is trying to achieve the same thing. A university team studying a poorly understood biological process might consider a new explanation or a promising research direction a major success. It doesn't necessarily need to lead to a commercial product. 

A company developing a new material, however, may need to know whether that material can be manufactured reliably, meets specific performance requirements or improves an existing product. Both kinds of research can benefit from AI, but they'll measure those benefits differently. 

An April 2026 working paper from the National Bureau of Economic Research looked at how AlphaFold2 had affected research into proteins. The researchers found that basic research involving previously less-studied proteins increased by 15% to 40% compared with proteins whose structures were already known. 

However, they hadn't yet found evidence of a similar increase in early-stage drug development involving those proteins. That doesn't mean the additional research wasn't useful. It simply shows that improving our understanding of something doesn't always lead to an immediate practical application. 

Sometimes the value of a scientific discovery takes years to become clear, which makes measuring the benefits of AI in research more complicated than looking at how much time it saves.

How Should Organisations Measure the Value of AI in R&D?

If an organisation is investing in AI to support scientific research, it needs to know what it's hoping to get out of that investment. Saving researchers time might be enough to justify using AI for certain tasks. 

But if the goal is to discover new materials, improve products or solve difficult scientific problems, the organisation needs to look at more than how quickly its researchers can work. The simplest place to start is by separating the different ways AI can improve research.

Measure productivity and scientific outcomes separately

There are four useful areas to look at when measuring the value of AI in R&D.

These measures are connected, but they tell us different things. For example, an organisation might find that AI has cut the time needed to analyse experimental data in half. That's a useful improvement. But it doesn't tell us whether the researchers are completing more experiments, finding better answers or reaching useful conclusions any sooner. 

To understand the wider benefit, the organisation would need to look at what happens after that analysis is complete.

• Are scientists using the time they've saved to investigate new ideas?
• Are they finding things they might otherwise have missed?
• And are those findings holding up when they're tested?

It's also worth remembering that some results take longer to measure than others. An organisation can see how much time AI saves almost immediately. But it might take months or years to know whether the research it helped produce leads to a useful discovery.

Evaluate AI against the goals of the research programme

Before introducing AI into a research project, organisations should understand how the work is currently being done and what's slowing it down. That gives them something to compare their results against once AI is introduced. It also helps them decide whether they're solving the right problem in the first place. 

For example, if researchers spend most of their time waiting for laboratory equipment, making their data analysis faster might not improve the overall project very much. But if they're struggling to work through large amounts of experimental data, AI could make a much bigger difference. 

There are a few questions organisations should be asking when measuring the results:

• Where is AI actually saving time? Is it helping scientists complete useful work faster, or creating more information they need to check?
• What are researchers doing with the time they've saved? Are they investigating new questions, improving existing work or running more experiments?
• Is AI creating extra work elsewhere? How much time and money are needed to check its results or test the ideas it produces?
• Are the results getting better? Is the organisation finding more reliable answers, making useful discoveries or moving closer to its research goals?
• How long will it take to see the benefits? Should the improvements be visible during the current project, or will they take longer to develop?

The answers can help organisations work out whether their AI investment is doing what they expected. They can also show where spending more money on AI might not be the best next step. If AI is already producing more possible discoveries than researchers can test, buying another tool to generate even more ideas probably won't help much. 

Investing in better testing facilities or giving researchers more time to check their results could be more useful. Organisations also need to account for the cost of checking AI-generated work. Producing more possible discoveries is only helpful if researchers have the time, equipment and resources to find out which ones are worth pursuing. 

For commercial R&D teams, these costs can form part of measuring AI research ROI, or the return they're getting on their investment. But not every useful scientific discovery can be measured in money, especially when the research is still in its early stages. A finding that helps scientists understand a disease might not produce an immediate financial return. 

It could still become an important part of developing a treatment years later. The best way to measure AI's value is to look at what the research was supposed to achieve, then work out whether AI is helping researchers get there.

Final Thoughts: Better Research Matters More Than Faster Research

AI is already helping scientists do things that once took considerably more time and effort. And as the technology improves, it's likely to become useful for more complex parts of scientific research. But there's still a difference between doing more research and making better discoveries. 

For organisations investing in scientific AI, the important question isn't simply how much time researchers are saving. It's whether those savings are helping them find more reliable answers, investigate worthwhile ideas and make progress towards their research goals. 

As AI takes on more scientific work, proving what those improvements actually achieve will become an even bigger part of deciding where to invest. And the same question applies well beyond scientific research. As businesses introduce AI into more of their work, understanding what they're getting in return is becoming just as important as understanding what the technology can do. 

EM360Tech continues to follow those developments, looking at how enterprise AI is changing the way organisations work and what they're actually achieving with it.