
The Most Important AI Question Has Nothing To Do With AI

What happens when critical technology becomes someone else’s decision?
A few days ago, something rather interesting happened. Not spectacular. Not cinematic. No robots escaped from laboratories, and no artificial superintelligence seized control of the internet while politely asking us to accept updated terms and conditions.
Instead, access to some of the most advanced AI models became a geopolitical issue.
According to Reuters, the European Commission is reviewing the practical consequences of a recent US export-control directive that required Anthropic to disable access to its most advanced models for foreign users. The details will matter, of course. They always do. But even before all details are settled, the moment itself is important.
For a brief moment, many Europeans were confronted with a surprisingly simple question:
What happens when critical technology becomes someone else’s decision?
Not because the technology stopped working. Not because users stopped paying. Not because the models became less useful overnight. But because access to them became subject to a decision made somewhere else, under priorities that may not be our own.
That is why I am not entirely convinced that AI is the real story here.
AI is the headline.
Dependency is the story.
The Anthropic Moment
Every generation seems to experience a moment when a technology stops being merely a useful tool and starts becoming infrastructure. At first, the technology is exciting. It attracts enthusiasts, sceptics, investors, consultants, policymakers, conference organisers, and the occasional person who has discovered a new three-letter acronym and is determined to use it before lunch.
Then something quieter happens. The technology becomes useful. Then it becomes normal. Then it becomes essential. Eventually, if it is truly successful, it becomes invisible.
Nobody spends much time thinking about electricity anymore. Most people do not wake up in the morning wondering how the power grid works. The same is true for the internet. Few people think about the networks, cables, protocols, data centres, and systems that quietly make modern life possible.
Successful technologies disappear from view.
And that is often the moment when they become most important.
GPS is a useful example. Most Europeans know how to use it. Far fewer know who operates it. Even fewer have seriously considered what would happen if access, accuracy, availability, or prioritisation changed. For most of us, GPS simply exists, like running water, street lighting, or the mysterious office printer that claims to be online but is clearly having a private crisis.
The same pattern emerged with cloud computing. At first, it was a technical discussion. Then it became a business discussion. Today, for many organisations, it is simply part of reality.
AI appears to be following the same path. And that should make us pause.
Every successful technology eventually becomes invisible. Every invisible technology eventually becomes infrastructure.
We Have Seen This Movie Before
Europe has spent much of the last decade rediscovering a rather old lesson: dependency is easy to ignore while everything works. Energy supplies were once treated as a technical and commercial matter. Semiconductors were once treated as a supply-chain matter. Cloud infrastructure was once treated as an IT procurement matter.
Then reality arrived, wearing heavy boots, and reminded us that infrastructure is never merely technical.
Infrastructure is strategic.
The interesting thing about dependency is that it rarely announces itself dramatically. It grows quietly. One useful tool becomes two. One convenient platform becomes ten. One practical shortcut becomes an operating model. Eventually, entire workflows, organisations, industries, and public services are built on technologies that were originally adopted because they were simply useful.
There is nothing inherently wrong with that. The world benefits from cooperation, trade, specialisation, and innovation. American technology, in particular, has produced extraordinary tools, and pretending otherwise would be both ungenerous and unserious.
The problem begins when convenience quietly replaces understanding.
Because understanding creates options.
Dependency reduces them.
That is why the current AI debate is so fascinating. Many people are asking whether Europe needs its own frontier models. Others ask whether Europe can compete with American or Chinese AI companies. These are important questions, but they may not be the most important ones.
Perhaps the deeper question is not whether Europe can access advanced AI.
Perhaps the deeper question is whether Europe can remain competent, resilient, and responsible when access changes.
The Productivity Puzzle
One of the strangest aspects of the AI debate is how quickly people became certain. According to some commentators, AI will transform productivity, reshape industries, and usher in a new era of prosperity. According to others, it is little more than a very expensive autocomplete system wrapped in marketing, venture capital, and suspiciously confident slide decks.
As is often the case, reality appears considerably less interested in taking sides.
The evidence is mixed, and that is precisely what makes it interesting. Research by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that a generative AI assistant increased productivity in customer support by around 14–15%, measured by issues resolved per hour. But the most important detail is not only the average gain. The study found that less experienced and lower-skilled workers benefited much more, while the impact on experienced and highly skilled workers was far smaller.
That should make us think.
It suggests that AI may not simply replace expertise. In many settings, AI seems to make parts of expertise more accessible. A junior employee often lacks structure, examples, vocabulary, patterns, and confidence. AI can provide some of that. An experienced professional already has much of it, so the immediate visible gain may naturally be smaller.
AI does not eliminate expertise. It changes how expertise travels.
As a financial modeller, I find this especially interesting. I do not need AI to explain discounted cash flow valuation to me. I do not need it to explain working capital, project finance, or why a balance sheet refusing to balance is the spreadsheet equivalent of a smoke alarm at 3 a.m.
But I may use AI to structure an article, challenge an argument, summarise a transcript, prepare training material, review a report, draft alternative explanations, or make a technical topic more accessible to a non-technical audience.
The role changes.
The value remains.
Easy To Use Is Not The Same As Easy To Use Well
This is where one of the largest misunderstandings begins. People say modern AI is intuitive, and they are right. Anyone can open ChatGPT, Claude, Gemini, or another language model and start asking questions within seconds. That is genuinely remarkable.
But there is a significant difference between being easy to access and being easy to use well.
Excel offers a useful comparison. It is easy to open. It is easy to type into a cell. It is even easy to produce something that looks, from a respectful distance, like a spreadsheet. That does not mean every spreadsheet is a robust, transparent, bankable financial model.
The interface is simple.
The professional judgement is not.
AI is increasingly similar. It is conversational, friendly, fast, and often impressive. But the real skill lies in knowing which task to give it, how to frame that task, how to check the answer, when to challenge it, when to ignore it, and when not to use it at all.
This may explain why AI often reduces productivity at the beginning. Many people expect an immediate gain. Instead, they receive a new kind of work. They must learn how to ask better questions, how to verify outputs, how to protect confidential data, how to avoid hallucinations, how to integrate AI into existing processes, and how to decide where human judgement must remain firmly in charge.
The Harvard Business School and BCG study on the “jagged technological frontier” captures this well. AI can improve performance on some knowledge-work tasks while worsening performance on others that look similarly difficult from the outside. That is a beautiful phrase, “jagged frontier”, because it describes what many practitioners experience every day. The border between useful and dangerous is not a clean line. It is uneven, surprising, and occasionally hiding behind a very confident paragraph.
AI is easy to operate. Professional AI use is a learned skill.
The Most Expensive Training Programme In History
There is another explanation for the apparent contradiction between impressive AI demos and disappointing AI rollouts. Perhaps AI is not primarily a productivity technology. Perhaps it is a learning technology that eventually produces productivity.
That may sound like a subtle distinction, but it changes the whole discussion.
Throughout history, major technologies rarely delivered their promised benefits immediately. Electricity did not transform factories the moment they connected to the grid. Computers did not revolutionise office work the day they arrived on desks. Enterprise software did not magically improve organisations simply because licences had been purchased and a consultant had used the word “implementation” with admirable confidence.
The technology arrived first.
The capability arrived later.
Brynjolfsson, Rock, and Syverson describe this through the idea of the productivity J-curve: general-purpose technologies such as AI often require major complementary investments in processes, business models, skills, and organisational change before their benefits appear in measured productivity. In plain English, companies spend money first, get confused for a while, redesign how they work, train people, make mistakes, and only later discover where the real value lives.
That sounds very familiar.
Many organisations currently possess AI. Far fewer possess AI competence.
And this distinction matters more than most AI sales presentations would like to admit.
Technology can be purchased. Competence has to be built.
Technology can be imported. Competence has to be developed. Technology can be licensed. Competence has to be practised, reviewed, challenged, and improved over time.
This also helps explain why some AI initiatives seem to disappoint. The MIT NANDA “GenAI Divide” report found that most enterprise AI initiatives studied were not producing measurable returns, with the core problem often linked not to the quality of the models but to integration and organisational learning. That is not a small detail. It suggests that the bottleneck may not be intelligence inside the model, but competence around the model.
Perhaps AI is not failing.
Perhaps many organisations are simply trying to harvest productivity before they have planted capability.
Europe’s Real AI Challenge
This brings us back to Europe. The popular question is whether Europe needs its own ChatGPT, its own Claude, its own frontier model. It is an important question. It may even be a necessary one.
But it is not enough.
A continent can have access to powerful AI and still remain dependent. It can build models and still lack the infrastructure, skills, governance, standards, and institutional confidence required to use them well. It can buy technology and still fail to build capability.
Europe’s real AI challenge is not only technological.
It is educational.
It is institutional.
It is organisational.
It is cultural.
Can Europe build frontier AI users? Can it develop AI literacy across schools, universities, companies, public administrations, and professional communities? Can it create standards that make AI outputs reviewable, explainable, and accountable? Can it support open-source alternatives, European infrastructure, data governance, and practical competence rather than merely celebrating the next impressive launch?
A continent does not become technologically sovereign because it owns technology.
It becomes technologically sovereign because it understands it.
Ownership can be acquired.
Understanding has to be built.
And understanding takes longer.
The Linux Lesson
This is one reason why I have become increasingly interested in Linux. Not because Linux is perfect. It is not. Not because commercial software is bad. Much of it is excellent. And not because I believe every public servant in Europe should be handed a terminal window on Monday morning and told that freedom begins with the command line. That would be brave. It would also be a memorable week for IT support.
Linux is interesting because it represents a different philosophy.
For decades, Linux was often chosen by people who valued understanding over convenience, transparency over black boxes, and control over dependency. It was never merely an operating system. It was also a way of thinking.
Today, many of the same questions are appearing in the AI debate.
Who understands the system?
Who can review it?
Who can challenge it?
Who controls the infrastructure?
Who remains accountable when something goes wrong?
These questions are not anti-technology. They are signs of technological maturity.
The moment we stop asking them is usually the moment we should become concerned.
Europe does not need to reject American AI. That would be unrealistic, unnecessary, and frankly rather silly. But Europe should avoid confusing access with sovereignty. Being allowed to use a tool is not the same as being able to rely on it under stress.
The Responsibility Problem
As the AI debate has evolved, I have become increasingly convinced that the greatest risk may not be that machines become more intelligent. The greater risk may be that humans gradually stop exercising their own judgement.
Not because they are forced to.
But because convenience is usually easier than responsibility.
Financial modellers learned this lesson long ago. A model is not trusted because it produces an answer. A model is trusted because the answer can be understood, challenged, reviewed, and explained.
Perhaps AI will eventually face the same test.
Not whether it can produce answers. It clearly can. Sometimes impressively. Sometimes dangerously. Sometimes with the confidence of a junior analyst who has just discovered EBITDA and may need to be supervised around enterprise value.
The real test is whether society can understand where those answers came from, whether they can be challenged, and whether someone remains accountable when they are wrong.
Because every important decision eventually raises the same question:
Who remains responsible?
Not the model.
Not the software.
Not the platform.
The human.
Always the human.
That principle becomes more important as technology becomes more capable, not less.
The Real Question
The great irony of the AI revolution may be that we have spent years asking whether machines will eventually think like humans, while a far more practical question has been waiting quietly in the background.
What should remain human, even in an AI-assisted world?
Perhaps the most important AI question has nothing to do with AI. Perhaps it is a much older question.
How much of our future are we willing to outsource?
To machines. To platforms. To companies. To other countries. And ultimately, to systems that may be useful, powerful, impressive, and still not responsible.
Technology creates opportunities.
Dependency creates risks.
History suggests that both tend to grow together.
Europe does not need less AI. Europe probably needs more AI: more experimentation, more innovation, more research, more entrepreneurship, and more serious adoption. But it also needs more understanding, more competence, more resilience, more standards, and more responsibility.
Because sovereignty is not the ability to use a technology on a good day.
Sovereignty is the ability to continue when that technology is unavailable, incorrect, or insufficient.
And perhaps that is the most important AI question of all.
Sources
- Reuters, EU Commission looking at practical consequences of Anthropic decision, 14 June 2026.
- Reuters, Anthropic disables top AI models after US foreign access order, 13 June 2026.
- Erik Brynjolfsson, Danielle Li, Lindsey Raymond, Generative AI at Work, NBER / Quarterly Journal of Economics.
- Fabrizio Dell’Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School / BCG.
- Erik Brynjolfsson, Daniel Rock, Chad Syverson, The Productivity J-Curve, American Economic Journal / NBER.
- MIT NANDA, The GenAI Divide: State of AI in Business 2025.
