Science,
with agency.

The UK’s AI sovereignty debate has focused heavily on who builds frontier language models. We think the greatest opportunity lies in what we do with them: applying AI at the scientific frontier to create new technologies, companies and economic strengths.

Where could the UK gain an advantage quickly by applying AI to problems in biology, materials, energy or other fields? Which discoveries could become industries in which Britain leads—and what would it take to turn that potential into economic value and greater strategic independence here?

Can AI help us notice when an entire field is building on shaky evidence? For years, researchers published studies linking particular genes to depression. Much larger studies later failed to support those links. Could agents have helped expose the weak statistics and repeated assumptions earlier?

AI-assisted review could trace the evidence behind a proposal, check whether the studies were large enough to support their conclusions and surface contradictory findings. Could this help researchers and funders identify work built on unreliable foundations and design more informative experiments? How do we do that while making room for unconventional ideas and keeping the agents’ reasoning open to challenge?

Could independent checks become a routine part of research? Agents could help rebuild software environments, run the supplied code, compare outputs with reported results and document what is missing. What would it take to offer this support through journals, funders and institutes, with less work for authors?

Recovering the same result from the same data is one test. Finding it again with new data or in another lab is another. How could AI help plan, coordinate and interpret those replications too? A successful rerun still leaves the scientific assumptions to examine; a failed replication needs explanation.

If AI reduces the work of reviewing proposals and administering grants, what new responsibilities could funders take on? Their view across programmes could help reveal shared bottlenecks, connect complementary teams and commission the datasets, methods or infrastructure that many projects need.

Could agents help maintain that overview, identify better collaborators and show when new evidence should change a programme’s direction? The question is how funding agencies could take on more of the coordination work that otherwise falls between individual grants, while giving researchers room to challenge the plan.

A paper captures only part of a developing scientific argument. What would change if questions, hypotheses, claims, evidence and disagreements were connected in a living record that people and agents could inspect, challenge and update?

Discourse graphs offer one starting point: making the reasoning behind a claim explicit and linking it to supporting or opposing evidence. How should review, credit and publication work when knowledge is built this way?

Electronic lab notebooks hold experiments, but the reasoning around them is often scattered across conversations, files and people’s memories. What would change if agents could use these records to track hypotheses, experimental plans, results and the reasons a lab changed direction?

Could they keep that account current, connect new results to the questions they bear on and help colleagues decide what to test next? The challenge is to make the lab’s scientific reasoning usable as work unfolds, with researchers able to inspect and correct it.

Research-service marketplaces and programmable cloud labs already exist. What would connect those capabilities into a market where people and agents can discover, compare and commission automated experiments across many providers?

An API needs a clear agreement about samples, protocols, price, turnaround and the data and quality checks returned. How do we make those services reliable across labs, help new providers get started and make access affordable? Which work belongs in shared facilities, and which could a distributed marketplace do better?

What would institutional intelligence look like in practice: a shared memory of what teams are trying to do, what they know, what capabilities they have and where progress is blocked? Think of a “company brain” for a research institute that stays connected to the work of its labs.

How could that help an institute identify common bottlenecks, connect complementary expertise and build the methods, datasets or collaborations a field needs? What changes to roles, incentives and decision-making would let the institute act on what it learns?