Four personal memories dramatise the pace of technological change that has happened in my lifetime. In 1983, aged 23, I bought a ZX Spectrum computer and wrote my first code on it. In 1995 I wrote my first web page and came to understand that it was the network, not the computer, that would now drive the information revolution.
In 2010 I first shot a video on an iPhone. And last month, I deployed the AI model Claude to create an agent simulating cascades of failure across the UK’s national infrastructure.
At each phase of the digital era I’ve been a dabbler rather than a techie: learning enough about coding to understand I could never do it; enough about networks to realise how they might decentralise capitalism.
But AI is an altogether different thing: its power depends on the imagination of the humans using it – not, ultimately, their skill. Purists may deride “vibe coding” – the creation of software by amateurs using plain text instructions – but AI brings the ability to “vibe” more or less anything digital into existence.
During each of these tech turning points, I remember being thrilled by the vistas of possibility that were suddenly opened. But nothing has been so simultaneously thrilling and alarming as the current revolution in agentic AI.
The rapid development of the US frontier models, together with the associated “stack” of services and regulations America is trying to impose as their condition of use, presents the UK with three simultaneous strategic dilemmas.
The first concerns security. Last week Anthropic, the owner of Claude, revealed that the model had broken out of a test environment – where it had been tasked with capturing a fictional piece of secret information – and done it for real against three organisations, one of which was a cybersecurity company.
Though the breakout was the result of human error, at no stage did the AI check with its human owner whether the attack was a good idea – and the “cleverest” version of Claude actually reasoned itself back into action after experiencing doubts.
Suggested Reading
Artificial intelligence and artificial philosophy
The UK hosts the global crown jewels of AI safety, in the AI Security Institute, which is one of the few non-US organisations trusted to test and evaluate the models of the Silicon Valley giants. Backed by No 10 and £66m of taxpayer cash, the AISI is the global first line of defence for businesses and civil society against potential harmful effects of AI.
And the obvious question is: what are we going to do to raise our defences against the latest models being deployed? AI is developing its capabilities so fast that the US government was already forced temporarily to suspend access to Claude’s advanced Mythos model, because of its power to crack cybersecurity.
But we are fast approaching a point where questions of safety come into conflict with the business imperatives of the providers, which demand the rapid rollout of ever more capable AI models. At what point does the British state – which has done so much to advance the UK’s participation in the AI revolution – do what Trump did with Mythos, if necessary to the detriment of its relationship with the firms and the US itself?
The second strategic dilemma concerns where the UK is going to sit within the geostrategic competition between the US and China over their visions for AI. The US models sit on vast US data centres, using chips that cannot be sold to China and making vast amounts of electricity unavailable to American consumers. Their stock market value depends entirely on the digital firewall erected around the core intellectual property – the design of the models themselves – and the vast amount of data they have been trained on.
Chinese AI is, by contrast, open source by design. You can run it on your own server, see the blueprint, tweak the design and its reasoning strategies – known as “weights”. China has chosen this model because, first, it has no access to the processors needed to run it the American way and, second, because Xi Jinping sees AI as a general purpose technology underpinning a new industrial revolution in business sectors that don’t currently exist.
The “New Quality Productive Forces”, in Chinese communist jargon, include the “low-altitude economy”, brain-machine interfaces and bio-engineering. Xi figures that if, instead of trying to monopolise AI, you provide it as a public good, you give these specialist Chinese industrial endeavours an unbeatable start in their competition with the west.
A year ago, the obvious downside of the Chinese model was that its products – Kimi, Qwen, DeepSeek and the like – were inferior to the US versions. Today however, the US versions are hitting up against the limits of what can safely be deployed, while the Chinese models are catching up fast.
But the US is fighting back. Trump has mused on banning Chinese open-source AI, or forcing businesses that use it to meet the insurance costs and give security guarantees.
For the UK, stuck in the middle, with no global-class AI model to join the race for supremacy, the choices are limited but stark. If Trump bans open-source AI, do we follow suit? Or do we join with European allies to create a “good enough” alternative. That is not something that can be answered case-by-case, but needs to be the product of geostrategic reasoning.
The appointment of Kanishka Narayan MP as a full cabinet minister in charge of AI is a welcome opportunity to bring that debate where it belongs – to the pubs and school gates of Britain rather than the rarefied seminars of the AI community.
The third dilemma concerns the potential transformation of the state. It is said that using an AI model available for $20 a month, civil servants were able to produce in 20 seconds an internal report that could have reasonably been allocated half a working week to a human. Though the quality was 85% of what a human could have produced, that discrepancy may not last very long.
Suggested Reading
Artificial intelligence and the end of capitalism
Right now, with professionals across the business world and civil service limited by corporate rules about which AI they can use, adoption is sluggish. Every office worker using Microsoft 365 now recognises the Copilot pop-ups that bring AI into the everyday tasks of report and email writing.
But these effects are puny compared with the massively transformative potential of AI agents. The task we should be setting AI is not “take over my calendar and prioritise my meetings”, but “redesign the Department of Work and Pensions to eradicate the millions of human hours wasted in meetings, assessments, reports and interviews, and make the experience of claiming benefits both painless and just”.
At the level of physical stuff, the UK has designated five AI transformation zones – in Lanarkshire, north and south Wales, Tyneside and Oxfordshire. But the constrained supply of electrical grid connections and inevitable local opposition means actual data centre plans in these areas are scarce.
So we are a long way from possessing a “sovereign AI” that could be unleashed safely into the civil service, local government or policing without creating massive dependency on the US. However, by embracing the goal of AI sovereignty, by creating the AISI and making London home to a growing number of AI firms, the UK is heading in the right direction.
On AI security, geostrategy and state transformation all the key decisions lie ahead of us. The challenge for politics is to subject as many of those decisions as possible to public scrutiny, before we get locked into pathways we might regret.
Most popular
Elon Musk’s The Odyssey: The Grok Imagine version is here, and it’s terrible
Help! Donald Trump keeps begging me for money!
The miserable triumph of Bernard Arnault, richest man in Europe
Trump has made his biggest, most dangerous mistake yet
