Data science and AI are already part of how cancer research is done. And yet the science may not be moving as fast as it could.

When we invest in the technology, it is easy to forget the foundations beneath it: the data we train models on, the trust and inclusivity around how it is used, the people with the skills to do the work, and the computing power to run it. These foundations will decide whether the UK’s strengths reach the people they are meant to help, or stall before they do.

The question is not whether we have the ambition. We clearly do, CRUK’s new Data Science in Cancer Research strategy is a good example of that. It is whether we can match that ambition with the harder, quieter work of building the foundations beneath it.

An exciting moment

Across the UK, AI is moving from promise to practice. It is already helping researchers read scans, design cancer vaccines and detect patterns across datasets too vast to analyse by hand.

The UK is well placed to build on this momentum: it is home to leading AI labs such as DeepMind, and Anthropic’s new London office will sit close to the Francis Crick Institute, strengthening the connection between cutting-edge AI and a world-class life sciences research base. Government investment is also increasing, from a planned twenty-fold expansion in public compute by 2030 to a new national supercomputer in Edinburgh and a Sovereign AI Fund that names life sciences as a priority. CRUK is backing the field directly too, investing £60 million through its Data Science in Research Strategy.

It is an exciting moment. But investment alone will not be enough. The quieter groundwork beneath these AI tools must keep pace.

An old problem with higher stakes

The barriers we set out are not new. Bringing the right data together, in a form researchers can readily use, has never been straightforward. People who can work fluently across data science and biology have always been in short supply, and useable computing power has often struggled to reach the researchers who need it.

What AI changes is the cost of leaving these things unfixed. A researcher could once work around a thin dataset or a clunky system through judgement and patience. A model cannot. It inherits the gaps in the data and infrastructure around it – and, when reused across multiple projects, can reproduce those flaws at scale.

Taken together, these foundational gaps mean that AI-driven research happens more slowly, works less well, or turns out less fair to the very patients whose trust the research depends on. That last point matters most, and it runs underneath everything else: trust is far easier to lose than to rebuild, especially in a climate where the public conversation about AI tends towards caution.

Four foundations to get right

None of this is cause for pessimism. With the right investment and support from government, these are solvable problems. The changes we are calling for fall into four broad themes.

The first is sorting out the data itself. The UK’s health data assets are world-class, but holding data and being able to use it are not the same thing. Too often a researcher with a good idea spends months not doing science, but chasing permissions across institutions, only to find the dataset they finally reach is missing the fields they needed. And even when data can be reached, it has to be in a form AI can work with: linked up rather than scattered, and held to shared standards rather than reinvented lab by lab. We see the forthcoming Health Data Research Service as a chance to build that in from the start, and cancer, with some of the most complex and sensitive data in the system, is the right test case. Get it right there, and it will work almost anywhere.

The second is making AI both inclusive and trustworthy. Who is in the data decides who benefits from the science. A model is only as good, and only as fair, as the data behind it, so the datasets we rely on need to reflect the whole population, not just the groups easiest to study. AI raises the stakes here, because it doesn’t only inherit the gaps in our data, it amplifies them. The answer isn’t to discard every imperfect dataset, but to be honest about what’s in them, so their limits are understood rather than discovered too late. Closely tied to this is trust: patients will only share sensitive data when they’re confident it will be handled with care, so we have to bring people into research decision making from the start, as partners rather than names informed after the fact.