Artificial intelligence has spent the past few years transforming how people write, search, programme and create.
The UK’s next ambition is considerably bigger.
The government wants artificial intelligence to become part of the machinery of scientific discovery itself, helping researchers discover medicines, design new materials, understand biology, advance quantum technologies and potentially accelerate the development of fusion energy.
Its AI for Science Strategy sets out an attempt to build the infrastructure required to make that happen.
The most eye-catching ambition is in medicine. By 2030, the government wants Britain to develop trial-ready drug candidates within 100 days.
That does not mean discovering a medicine on Monday and giving it to patients three months later. Clinical trials, regulatory approval, manufacturing and safety monitoring would still be required.
Instead, the ambition is to dramatically compress the early drug-discovery process that happens before human trials begin.
If Britain can achieve that safely, the implications could extend far beyond pharmaceuticals.
It would demonstrate that artificial intelligence can move from helping scientists analyse knowledge to actively accelerating the creation of new scientific knowledge.
AI Is Moving From Prediction to Discovery
One of the defining breakthroughs in scientific AI came from understanding proteins.
Proteins are fundamental components of life, but their function depends heavily on their three-dimensional structure. Determining those structures experimentally can be difficult and time-consuming.
AI changed the problem.
Systems such as AlphaFold demonstrated that machine learning could predict protein structures with extraordinary effectiveness.
But the frontier has already moved again.
Scientists are increasingly using AI not simply to predict structures that exist in nature but to design entirely new proteins with particular properties.
That distinction is profound.
Prediction asks the computer to understand something that already exists.
Design asks it to create something that has never existed before.
The UK government’s strategy describes this transition as AI moving from prediction towards action.
That is where AI for science becomes particularly interesting.
What Does a 100-Day Drug Actually Mean?
Traditional drug development can take more than a decade from early discovery through clinical trials and eventual regulatory approval.
A considerable amount of that time is spent before a potential medicine ever reaches a patient.
Scientists need to understand a disease, identify a biological target, find molecules capable of interacting with that target, optimise promising candidates and determine whether those compounds have characteristics that make them suitable for further development.
AI could potentially accelerate several of these stages.
Models can search enormous collections of chemical structures. They can analyse scientific literature and biological datasets. They can help predict whether molecules will bind to particular proteins. They can identify potential toxicity problems earlier and help researchers decide which experiments are worth conducting.
The government’s Mission One therefore concentrates heavily on the preclinical discovery process.
The ambition is to combine AI, high-quality scientific data, large-scale computing and automated experimentation so that promising drug candidates can reach trial-ready status dramatically faster.
The target is 100 days by 2030.
That is an extraordinarily ambitious objective.
It is also a useful example of how AI could change science.
The Scientist Could Eventually Have an AI Research Team
Generative AI is usually discussed as a tool that responds to human instructions.
Scientific AI could become considerably more autonomous.
The government’s strategy anticipates increasingly capable AI science agents that can search scientific literature, analyse datasets, generate hypotheses and assist with experimental design.
Imagine a researcher investigating a new treatment.
An AI agent could examine thousands of scientific papers and datasets that no individual researcher could reasonably read. It could identify previously overlooked relationships, propose potential biological targets and suggest experiments capable of testing those hypotheses.
Another model could examine possible molecules.
Another could predict toxicity.
Another could analyse experimental results.
The human scientist would remain responsible for scientific judgement, validation and interpretation, but the amount of intellectual territory that could be explored simultaneously would expand enormously.
AI could effectively increase the number of hypotheses science is capable of testing.
Then Connect the AI to a Robot
The next step is even more significant.
Instead of AI merely recommending an experiment, scientists can connect AI systems to automated laboratories capable of conducting experiments physically.
This concept is sometimes called a self-driving or autonomous laboratory.
The UK’s Materials Innovation Factory at the University of Liverpool has already demonstrated part of this future. Researchers developed a mobile robotic chemist capable of performing experiments largely autonomously.
The robot conducted 688 experiments over eight days while searching for an improved catalyst.
This creates a feedback loop.
The AI proposes an experiment. The automated laboratory performs it. Sensors record the result. The system analyses what happened and decides which experiment would be most informative next.
Scientific discovery starts becoming iterative at machine speed.
The laboratory does not sleep.
Failed Experiments Could Become Valuable Data
There is another fascinating part of the strategy that receives considerably less attention.
Science produces enormous quantities of data that never become easily accessible.
Successful experiments are published. Interesting results appear in journals. Failed experiments, negative results and abandoned approaches frequently remain inside laboratory notebooks, university servers and research databases.
This is sometimes described as scientific “dark data”.
For human researchers, reading millions of failed experiments would be practically impossible.
For AI, those failures could be extraordinarily valuable.
Knowing that an experiment did not work prevents another researcher from repeating the same mistake. More importantly, large collections of negative results could help machine-learning systems understand the boundaries between what works and what does not.
AI therefore changes the economics of failure.
An unsuccessful experiment may still produce valuable training information.
The UK’s strategy specifically recognises the importance of improving scientific data infrastructure, including capturing negative results and other underused information that could support AI-enabled research.
The scientific record of the future may therefore need to contain considerably more than published successes.
Britain Is Building a Health Data Infrastructure
Medical AI requires something else that Britain possesses in unusual quantities: health data.
The government and Wellcome are backing a new Health Data Research Service with investment of up to £600 million.
The service is intended to provide approved researchers with a secure, single point of access to health datasets drawn from multiple sources.
Its importance to AI could be substantial.
Imagine securely connecting appropriately governed datasets covering pathology, radiology, genomics and clinical information.
Researchers could use these resources to identify biomarkers, understand disease progression, improve clinical-trial design and build models that better predict how people might respond to treatments.
The Health Data Research Service is expected to begin delivering capabilities during 2026 and progressively expand towards full operation by 2030.
The objective is not simply to create another government database.
It is to create research infrastructure.
Why Data May Matter More Than the AI Model
The public conversation about artificial intelligence tends to focus on models.
Science may expose why that is only part of the story.
An AI system trying to discover a drug is constrained by the scientific information available to it.
If the underlying data is incomplete, inconsistent or poorly structured, making the model larger does not necessarily solve the problem.
This is why institutions such as UK Biobank and EMBL-EBI could become strategically important to Britain’s AI ambitions.
The UK also has emerging projects such as OpenBind, which is generating large quantities of experimental information about how potential drugs interact with proteins.
The government has invested £8 million in the OpenBind consortium at Diamond Light Source with the ambition of creating an exceptionally large collection of high-quality protein-ligand interaction data.
That data can then help scientists build better AI models for early-stage drug design.
The competitive advantage in scientific AI may therefore belong not simply to the country with the largest model.
It may belong to the country with the best combination of models, data, computing infrastructure, laboratories and scientific expertise.
Five Frontiers Britain Wants to Target
The strategy identifies five priority areas where the government believes Britain has scientific strengths and AI could accelerate discovery.
| Frontier | What AI Could Help Scientists Do |
|---|---|
| Medical research | Discover drug targets, design molecules, predict biological responses and accelerate clinical research. |
| Advanced materials | Search enormous combinations of materials to discover new properties for energy, manufacturing and industry. |
| Engineering biology | Design biological systems, proteins and organisms for medicines, chemicals, materials and fuels. |
| Quantum technologies | Assist with system design, optimisation and the development of future quantum technologies. |
| Fusion energy | Improve plasma control, materials discovery, simulation and future fusion-plant operations. |
These fields appear different, but they share a common problem.
The number of possible experiments is enormous.
Scientists cannot test all of them.
AI can help decide where to look.
Science Needs Different People Too
The strategy also recognises something frequently overlooked in discussions about artificial intelligence.
AI-enabled science requires more than AI researchers.
It needs scientists who understand machine learning, but it also needs research software engineers, data specialists, laboratory automation experts, technicians and people capable of connecting computational systems with physical experiments.
The government wants to significantly expand the pipeline of researchers capable of working across these boundaries, including through interdisciplinary fellowships, doctoral training and new career pathways.
This matters because some of the most important scientific breakthroughs may increasingly occur between disciplines.
A biologist may understand the disease.
A machine-learning researcher may understand the model.
A software engineer may build the infrastructure.
A robotics specialist may automate the laboratory.
A data engineer may make the experimental information usable.
The breakthrough happens when those people can work as one research system.
Compute Becomes Scientific Infrastructure
There is another resource required for this transformation: computing power.
Advanced AI models require enormous computational resources, and individual research groups cannot necessarily afford the infrastructure needed to train or operate them.
The government therefore plans to provide researchers with access to Britain’s national AI Research Resource.
Some projects will be able to request hundreds of thousands of GPU hours over several months.
For particularly ambitious scientific missions, teams will be able to apply for temporary access to an extraordinary proportion of Britain’s national AI supercomputing capacity.
That represents an important conceptual change.
Supercomputing is becoming part of the laboratory.
A modern scientist may require microscopes, experimental equipment and laboratory space, but increasingly they may also require thousands of GPUs.
The £2 Billion Needs Some Explanation
The numbers surrounding the strategy can easily become confusing.
The government is investing £2 billion between 2026 and 2030 to deliver the wider AI Opportunities Action Plan.
The AI for Science Strategy itself directs up to £137 million of that investment towards its specific programme of scientific AI activity.
That sits alongside much larger investments elsewhere in Britain’s research infrastructure, including supercomputing and the Health Data Research Service.
The £600 million associated with the Health Data Research Service is also not entirely government funding. It is backed jointly by the UK government and Wellcome.
These distinctions matter.
Britain is not simply placing £2 billion into a pot labelled “AI for Science”.
It is constructing several pieces of infrastructure that are intended to work together.
The Bigger Picture
The first wave of generative AI changed how people interact with information.
AI for science could change how humanity creates new information about the physical world.
That is a considerably more difficult challenge.
A chatbot can generate an incorrect sentence and somebody can correct it.
A scientific AI proposing a drug, chemical or material operates in a world where mistakes have physical consequences.
Predictions therefore need experiments. Models need validation. Data needs provenance. Autonomous laboratories need safeguards. Scientific conclusions still need reproducibility.
The government itself acknowledges that fully automated scientific discovery remains some distance away.
That caution is important.
But the direction is becoming clearer.
AI systems are moving from analysing scientific knowledge towards generating hypotheses, designing molecules, proposing experiments and interacting with automated laboratories.
Britain’s strategy is effectively a bet that the countries capable of connecting these systems to high-quality data, enormous computing resources and world-class scientists will have an advantage in the next era of discovery.
The 100-day drug mission provides a useful test.
If AI can help scientists compress years of early drug discovery into months without sacrificing scientific rigour or safety, the consequences will reach far beyond pharmaceuticals.
It could change what we mean by the speed of science itself.
The next great AI race may not be about which country builds the smartest chatbot. It may be about which country builds the scientific infrastructure that allows AI to discover something humanity did not know before.