Blog
AI and science are colliding. How will they fuse?
2 March 2026
Artificial intelligence is increasingly permeating the frontier of scientific research and development. The exponential growth of AI related papers in recent years, and the diffusion of AI from mathematics and computer science to other disciplines present a challenge common to emergent phenomena. Potentially transformative effects of AI are playing out rapidly, yet the nascence of the technology means that there has been relatively little time to assess its impacts.
AI in science has become an important topic to science policymakers, funders, researchers and innovators who hold hopes that it will accelerate scientific discovery and support new breakthroughs. However, this rapid ascent triggers concerns that the technology will focus scientists on ‘AI-shaped’ problems at the expense of other areas, leading to diminishing returns and a lack of research diversity. Others acknowledge that while AI predictions can be useful, they do not necessarily increase our understanding of natural phenomena and their origins.
It is critical that we gain a better understanding of the ways in which AI tools are adopted in research and the impact they have, to ensure that AI integrates with and enhances science.
AlphaFold 2 – A case study
AlphaFold 2 is an important milestone in the development of AI for science. Released by Google DeepMind in 2021, the tool addresses the longstanding challenge of predicting protein structures. Identifying protein structures is a key component of investigating many biological mechanisms and enabling R&D, such as drug discovery. A solution to the protein structure prediction problem had evaded computational biologists for decades, but AlphaFold 2 represented a breakthrough, effectively predicting the vast majority of rigid protein structures and earning its creators the 2024 Nobel Prize for Chemistry.
The performance of AlphaFold 2 and its sudden, disruptive debut in the scientific ecosystem make it a useful case study to investigate how AI impacts and integrates with science. In a recent study from IGL’s Data and Technology Unit, we investigated the reach of AlphaFold 2 and its impact on scientific productivity, discovery, and applied innovation.
In this blog we aim to share some broader and near term implications of our key results which may be useful for those thinking about optimising the use of AI in science or further studying it.
Machines for discovery at scale
AlphaFold 2 is a tool that does a very good job on a specific type of problem: one that represents a bottleneck for a field, involves a large possible solution space, and possesses training data for accurate modelling. The category of AI that addresses this kind of problem could be described as ‘narrow’ AI, in contrast to general AI tools such as large language models (which are also seeing significant uptake in science).
One of our findings is that 5 years after its release, the reach of AlphaFold 2, as measured by direct and indirect citations, is expanding, and that this expansion is accelerating. This is in contrast to other high impact contemporary biology research, where the growth in reach tapers off. We also found that the papers building on AlphaFold 2 disproportionately resulted in impactful science, as measured by citation metrics.
This suggests two things. First, narrow AI tools which solve critical problems can quickly gain widespread traction within their field (and beyond). Second, it is therefore not unreasonable to think that such tools will be instrumental in significant discoveries in high dimensional solution spaces that otherwise might have taken much longer to achieve, and which will be seen as breakthroughs in decades to come.
Making predictions and doing science are different but complementary
When a protein structure is predicted, it needs to be experimentally validated. This step has historically been expensive, long and laborious, incentivising researchers to focus on target structures where results are reasonably assured and worthwhile. However, our study suggests AlphaFold2 disrupts this risk aversion. We found that researchers building on the tool were significantly more likely to publish and validate new protein structures. They are also more likely to submit proteins with structures that are dissimilar to ones which are already known. In other words, AlphaFold2 is helping researchers determine more structures from less charted parts of the protein universe.
This highlights a broader potential: if AI lowers the risk of investigating less-explored parts of a knowledge space, it might be able to accelerate progress towards novel discoveries and inventions. For example, an AI trained to predict the properties of a material based on chemical composition could be used to screen millions of generated candidates for room temperature superconductors. It might suggest unconventional material designs that might have previously been overlooked, but which could then be tested in a lab, with researchers having greater confidence in obtaining the desired results.
Critics have reasonably pointed out that this kind of work is not science in the fullest sense, and more akin to engineering. It might provide a boost for innovation, but is less good for research that seeks to deepen our understanding of natural phenomena. This however, overlooks the fact that applied discoveries can be a precursor for new knowledge. Graphene is a prime example of this. The ability to produce two-dimensional materials opened possibilities for the creation of new materials and electronic devices and it also served as a catalyst for fundamental physics research.
Integration is the bottleneck for AI
For transformative effects, you likely need high integration between existing processes and AI use, or a transformation of existing processes altogether. Our study suggests some degree of integration between AlphaFold 2 and traditional research methods – we observe positive impacts of the tool on exploration, and a higher rate of experimental protein structure validation. This is at odds with the idea of the two acting like oil and water, and has likely been helped by AlphaFold 2 being free, accessible and accurate.
However, we also found that research linked to AlphaFold 2 had only a moderate and mixed impact on applied innovation outcomes. We looked at patents, clinical articles, and disease related research, and while we observed benefits associated with work built on AlphaFold 2, we often saw other frontier developments in structural biology having a similar level of impact.
Why did we not see a large and obvious boost in innovation downstream of AlphaFold 2? One reason might be that a five year time window is not long enough to observe applications of the tool. Another factor at play may be that AlphaFold 2 addresses only a narrow slice of R&D activities involved in downstream innovation. It does not, for example, provide an end-to-end, integrated solution for drug discovery or protein design.
Our study leaves room to speculate that systemic and transformative effects of AI will only be realised with some further developments. One is more flexible AI tools that perform well on multiple related, narrow tasks within a problem area. The other is the creation of applications that wrap around AI models to tightly integrate them within additional R&D workflows, including automation of laboratory work. Google DeepMind’s development of AlphaFold 3, and new products such as the one being created by materials science AI startup, Polaron, show how these two issues are already being recognised and addressed.
Outlook
Our findings provide a certain level of confidence in the potential for transformative impacts of AI in science. There is a strong potential for AI to unlock progress across a range of science and innovation frontiers, particularly if positive feedback loops between current approaches to science and AI-powered science are encouraged, and if integration is improved through iterative enhancements to tools and workflows. The creation of those conditions is not a given, and risks such as the creation of AI monocultures are real. At present, we should be wary of terms such as “revolution”, when describing the impact of AI on science. Harnessing AI for science in a way that maximises impacts, while avoiding pitfalls, will rely on humans navigating new, open waters, balancing permissive experimentation and stewardship of the scientific ecosystem from policymakers, funders and researchers themselves.
This is Part 1 of a 2 part series. The next article will discuss the potential for future work in this area, as well as some of the existing policy efforts to achieve a positive transition for AI in science.