Oxford and UCL just became central to the UK's AI transformation, and the Government's £60 million bet shows sovereignty is being taken seriously. However, responsible AI is the only way to make this successful.
Zoya Yasmine / Better Images of AI / CC BY 4.0
I'll be candid, and unlike some of the political elite, will declare my interests upfront. I did my undergrad at UCL and my postgrad at Oxford. So, when the UK Government announced last week that it plans to spend £60 million on two new research labs at both institutions, I felt somewhat chuffed. Both as an alumnus, but also in a professional capacity, as someone passionate about helping to build a future of safe and responsible AI.
Beyond my personal sentiments, this news is one of the most strategically interesting moves in the tech space that the UK has made this year, and deserves more attention than it is currently getting.
Beyond the headline
The polished press release aside, we need to get into the weeds of what this means for research and commercial application moving forward. So here is what the labs will actually be doing:
The Science of Fundamental AI Research (SOFAIR) lab at UCL will be tasked with developing open-source AI that can run on widely available hardware. This opens up applications of advanced AI on both ordinary consumer devices, and controllable, interrogable and ring-fenced machines.
The British Open-ended Learning and Discovery (BOLD) lab at Oxford will lead in reframing how AI learns about the world, aiming to find more efficient and adaptive ways to train models without the need for vast quantities of centralised compute capacity.
The undercurrent of this decision is a strategic departure from trying to keep up with the AI race led by US-based frontier labs and hyperscalers (as if we had any hope in the first place).
Rather, it signals the irrelevancy of that very race.
Shifting the narrative around the compute arms race
With US tech companies spending billions on infrastructure and compute to power AI, the UK has no hope of keeping up with that level of investment. But that doesn't mean we will lose. It is misguided to think that the current AI growth flywheel of bigger models, more compute, and more investment is the only option. In reality, that is likely the most irresponsible option when considering the social, environmental, political and economic consequences of unchecked AI development.
What SOFAIR and BOLD point to is a different, uniquely novel approach.
The next major breakthroughs in AI may not necessarily come from more compute power. Rather, they will emerge from the more targeted application of that power. When we rethink how to architect AI around smaller, more controllable models, AI that actually adds value can become cheaper, easier to adopt and less dependent on US Big Tech. These are important considerations when we ask questions about AI sovereignty.
Should our hospitals, public services, banks or councils be dependent on opaque frontier models based in the US?
These views are not just theoretical or academic. NVIDIA (the prevailing power station behind current AI infrastructural development) researchers have published a paper arguing that smaller models are powerful enough to deliver real world value in agentic use cases. When the company that is selling compute says that you don't always need more compute, we ought to take that as a meaningful signal about the future evolution of practical AI innovations. SOFAIR and BOLD are aligned to this trend, but are we ready to accept that trade-off?
Sovereignty is about choice over control
Looking beyond the buzzword that has been largely devoid of tangible meaning over recent years, recent developments in the practical implementation of sovereignty have made it a worthwhile attention grabber.
The key to unlocking successful sovereign AI is choice. The ability to choose models, to choose infrastructure, to choose data, to choose governance, to choose to change models, and importantly, to choose the level of control you have over those things. It's about independence from US Big Tech, without isolation from it. This sliding scale, choosing the right level of sovereignty based on the criticality of certain workloads, is precisely how we approach sovereign AI with our customers at Kainos.
Up until now, we have had little choice. To do AI seriously meant to depend on US Big Tech, be at the mercy of it, and always come second to US national and economic interests. If we could trust US labs wholeheartedly, conversations around sovereign AI would be redundant. However, past experiences have proven otherwise, not just in regard to AI, but also privacy, data misuse, misinformation and online harms. We've all been victims to decades of unjust socioeconomic externalities that stem from the capitalistic growth imperatives of US Big Tech, where the most adversely impacted tend to be the most marginalised and vulnerable in society. This has fundamentally eroded trust. As such, the need for sovereign AI is paramount.
SOFAIR and BOLD act as meaningful pivots away from our over-reliance on hyperscaler compute, helping make sovereign AI a practical reality. There is still a long way to go – we cannot forget about responsible governance, assurance, commercial vehicles, skills, and partnerships – but this is a step in the right direction. Though, without responsible AI, these efforts will be in vain.
Seeing the bigger picture
Zooming out, this investment in UK sovereign AI seems to be the missing piece in our broader UK AI strategy. It aligns with the £500m Sovereign AI Unit backing British AI startups, the AI Security Institute, the UKRI's AI Research Resource, and broader industry-leading organisations like the Alan Turing Institute and Ada Lovelace Institute. These compiled together provide the foundations of an interconnected, comprehensive and holistic national AI strategy.
If executed well, the ramifications of this will be bigger than the headline funding (which is, in reality, minuscule compared to the gargantuan investments US-based frontier labs receive). If the ecosystem works as planned, UK doctoral students who would have drifted to the Bay Area get a credible reason to stay. UK start-ups get a feedstock of fundamental research they can build on without licensing it from someone else. UK regulators get domestic experts who can actually advise them. UK public services get models they can adopt without nervously checking the small print of an American sub-processor agreement.
The success of this, as aforementioned, relies on its execution. The only way to ensure successful execution, is to ensure responsible AI is baked into our national AI strategy.
For this sovereign AI endeavour to be successful, it needs responsible AI.
As outlined in The Sovereign AI Unit's fifth priority frontier (which should be first in my opinion):
'The UK does not need to build every model. It can define the conditions under which models are trusted.'
Those two sentences should be on every AI policy whiteboard in Westminster, for they epitomise, in the cleanest articulation I've seen, where the UK's actual AI advantage lies.
The responsibility trade-off with sovereign AI
Open-source AI, smaller models, decentralised training, novel learning paradigms, they are all brilliant for accessibility, productivity and sovereignty. They are also, if we're honest, brilliant for increasing risk. Open-weight models are harder to recall when something goes wrong. Smaller models running on consumer hardware are harder to monitor centrally. Novel learning paradigms (or open-ended learning) purposefully deviate from controls originally baked in. The whole point of our AI strategy is novelty, which by definition opens us up to a whole world of novel risks.
We've been swimming against the tide over the last few years trying to get organisations to take responsible AI seriously. One of the worst things we could do is treat 'British' and 'open-source' as synonyms for 'safe by default'. They are not.
If SOFAIR and BOLD are to be the new foundations of our sovereign AI strategy, responsible AI needs to be the bedrock on which they are built.
Open by default needs to be governed by design. Open-weight models still require provenance, evaluation, red-teaming and clear documentation. Open source is not an excuse to skip assurance, it's an obligation to make assurance public.
Human-centred approaches need to be baked in from the lab bench. 'Practical, human-centred AI' is in BOLD's own framing. But this promise needs to be delivered, with social scientists and qualitative researchers forming part of the team. The communities AI impacts most – NHS patients, benefits claimants, small businesses, and everyone in-between – also need to be in the room.
Sovereign-by-design supersedes sovereign-by-accident. Sovereignty has to be engineered in across the AI lifecycle. The labs should be working with industry, policy and academic leaders on a shared assurance backbone, ideally mapped onto a comprehensive responsible AI and AI governance framework, so that outputs are trustworthy, scalable and defensible.
AI talent needs to reflect the country it serves. Ten doctoral students per lab is a start, but if the next generation of UK AI researchers lacks diversity, all looking and sounding the same, coming from the same places, we've recreated the very concentration problem we're trying to escape when it comes to fair, diverse and equitable AI outcomes.
If we get this right, Britain doesn't just have two new AI labs. Rather, we have a credible national AI stack across research, models, infrastructure, responsible-by-design assurance and talent, to keep the flywheel of sovereign UK AI spinning.
If we get it wrong, we'll have spent £60 million on producing efficient ways for irresponsible AI to become the norm. An overly technocentric execution of our sovereign AI strategy will lead to its fundamental failure.
The important questions
Britain's competitive edge in AI was never going to be compute. It was always going to be the combination of world-class research, mature institutions, and a regulatory culture serious enough to be trusted (it turns out that having a history of more than 250 years can also help in writing the next chapter of our AI story), with broader ramifications beyond our own shores. SOFAIR and BOLD are a meaningful, material vote in favour of leaning into that edge.
However, sovereign labs do not guarantee sovereign AI. Decisions do.
We ought to ask questions about openness, assurance, trust, who gets to use what and why, and on whose terms. Sovereign AI will reframe the power dynamics of the UK AI industry. But we need to ask delicate questions about who controls this power, and understand the nuances of who can take it away.
We can clearly build a sovereign AI ecosystem in the UK. So, the question I'd put back to my old universities, to UKRI, to the Government, and to anyone reading this:
Can we build sovereign AI in a way the public can trust?
Trust is earned, not announced in a press release. Let's make sure we prove it.
Let's not get lost in the pursuit of relentless innovation that leads nowhere.
Let's ask the uncomfortable questions.
Let's debate.
Only then can we build a future of sovereign, safe, and responsible AI in the UK.