Edition 31: AI Is Becoming Your Fifth Friend. Who’s Raising It?

Group Of Humans, Technology Partner, Josh Klein advises governments on AI sovereignty. His conclusion is blunt: building a national LLM is the wrong goal. Real sovereignty, he argues, is about controlling data, culture and alignment in a world where AI is becoming part of how citizens think and behave.

When governments talk about sovereign AI, infrastructure is usually the first point of focus. Josh Klein thinks that misses the point.

“I’ve been advising governments on sovereign AI for three years, and the conversation usually starts in the wrong place – infrastructure, data centres, national models. Those matter, but the more urgent question is different: who shapes your citizens’ behaviour? Data centres are a five-year play. The AI influencing how people think and act is happening now.”

To put it more simply, “If you’re the average of your five closest friends, AI is about to become one of them. That’s 20% of someone’s cognitive input. For governments, that’s an unprecedented form of internal soft power – not censorship, not propaganda, but a way to shape how citizens engage with public services, preserve cultural values, and think about civic life. The question is: whose values get embedded?”

In Klein’s view, AI creates a form of internal soft power that governments have never had before. Unlike social media, which has fragmented across platforms, large language models (LLMs) are becoming a single major interface between citizens and information.

“What’s different about LLMs versus social media is consolidation. Social media fragmented across dozens of platforms – no single actor owned the whole picture. LLMs centralise the interface. If citizen-government interaction runs through an AI layer, you get both influence and insight. Every query tells you something about what people need, what frustrates them, what they’re trying to accomplish.”

This, he argues, is where the real power of sovereign AI lies: breaking data silos, adding a semantic layer across government systems, and finally giving the public sector something like a P&L.

“Imagine AI handling citizen interactions across every ministry – licensing, benefits, healthcare, education. You’d instantly know: What are people struggling with? Where are the bottlenecks? Which services generate frustration versus satisfaction? Governments have never had that kind of real-time feedback loop. The potential for improving service delivery – and for understanding what citizens actually need – is enormous.”

“It cuts right to the heart of the issue governments have around growth, efficiency and reduced costs. It has massive transformational potential.”

But there’s a challenge on the sovereignty side. “At the same time governments are recognising this has an enormous potential impact, they’re also realising that, ‘oh by the way… a US company is sucking in all of this data and then producing its own results - and it’s a black box.’ And given the current geopolitical situation, that’s not appealing to anybody, really.”

Because, Klein says, that idea is already obsolete.

“The ‘build a national LLM’ strategy is already showing its limits. Models advance too fast – there’s a new frontier every few months. Your citizens are using Claude and GPT-4 today; you can’t hand them a three-year-old model and call it sovereignty. Real sovereignty is architectural: keep sensitive data in national or regional stores, use the best available models for your needs, and maintain the flexibility to switch when capabilities improve or when geopolitics shift.”

That forces a redefinition of what sovereignty actually means.

“Sovereignty isn’t just data residency – it’s inculturation. How do you make an AI that doesn’t default to Northern Californian assumptions, but actually reflects local language, social norms, and ethical frameworks? In Iceland [where Klein lives], linguistic fidelity matters enormously – there are only 370,000 Icelandic speakers. If AI erodes that language, you’ve lost something irreplaceable. Every country has these cultural imperatives and the framework has to accommodate them.”

In practice, this means:

• Making sure sensitive data is kept in national or regional stores to provide data privacy protections.

• Using external models, but making sure results are culturally, legally and tangibly localised in a way that makes them meaningful and recognisable.

• Designing systems so governments can switch models when geopolitics or company values shift in a way that no longer aligns with their own.

“The architecture has to be model-agnostic. In Iceland, we’re coordinating both Google and Anthropic under the same governance framework – if one vendor’s direction diverges from government values, or if a better option emerges, you can switch. This approach is rare yet validates the model, and is really the only sane approach when capabilities advance monthly and geopolitics can shift overnight. The goal is immediate value for ministries while maintaining strategic flexibility.”

Sovereignty, in this framing, is strategic optionality rather than ownership of a single national stack.

“You might query a frontier model hosted elsewhere, but your sensitive data stays local. You can tokenize, aggregate, use local models as intermediaries for certain functions. The key is optionality: if a provider’s values or a geopolitical situation changes, you’re not locked in. You can move.”

“Each nation needs to figure out on a case-by-case basis where it can exploit particular opportunities, because every country is different and has a different resource base.”

In places with high mobile penetration but different infrastructure profiles, Klein sees alternative approaches emerging.

“In places with high mobile penetration but limited traditional infrastructure – parts of East Africa, island nations, rapidly developing economies – the Western data-centre model doesn’t necessarily fit. What I’m exploring for contexts like these is federated learning: push computation to mobile devices, aggregate results centrally, iterate. It’s a different sovereignty architecture that works with existing connectivity patterns rather than requiring massive capital expenditure upfront. The leapfrogging potential is significant.”

When asked whether current government initiatives are genuine or mostly symbolic, he’s measured but direct.

“Right now, a lot of government AI announcements are aspirational – political signaling ahead of implementation. ‘We saved X million with AI’ often means someone projected it in a slide deck. Real capability shows up in sustained usage, ministry-level integration and measurable changes in how government actually operates. That’s what separates announcements from outcomes.

“The mistake is treating AI like previous digital transformations – cloud computing, e-commerce – where you’re digitising and scaling existing processes. AI is categorically different. It doesn’t just automate what you were already doing, it changes how people think, how they access information, how they make decisions. That’s why the governance questions are so much harder.”

The implication is uncomfortable but clear: sovereign AI is not primarily a technology project. It’s a cultural, behavioural and governance project.

“People focus on AI making mistakes – hallucinations, errors. But humans make mistakes constantly; we just call them cognitive biases or institutional failures. The question isn’t whether the system is perfect. It’s whether the system is aligned with your values, whether it’s resilient when things go wrong, and ultimately, who designed it and whose interests does it serve?”

“The real competition isn’t over servers and data centres – though those do matter – it’s over the cognitive infrastructure of your citizens: who shapes how they access information, how they interact with government, how they think about their place in society? That’s what sovereign AI is actually about.”