I’m very European. I believe in European values, and I think the European Union is one of the most amazing institutions and projects the world has ever seen. A continent that spent centuries at war chose cooperation. I don’t take that for granted.
But I worry about how we talk about AI.
Whenever the subject comes up, we reach for the high ground. We will build the most human AI. Not like China, the dictatorship. Not like the United States, the Wild West. What that more human AI actually means, nobody seems to know.
I would love a more human AI too. The problem is the order. That mission cannot come in front of us becoming a competitor. Values without power are a lecture, not a project.
And the power is slipping. According to the Stanford AI Index, US-based institutions produced 40 notable AI models in 2024. China produced 15. Europe, three. Private AI investment in the US reached $109 billion in 2024, nearly 12 times China’s. We are very fast becoming irrelevant on the global stage.
A union that can’t compete won’t get to set the terms, human or otherwise. I say this as someone who loves the project, and that’s exactly why I don’t want us to mistake the mission for the result.
I grew up hearing that the euro would make Europe a single market. One currency, one bloc, 450 million consumers. The euro did away with exchange rates. It did not do away with the borders.
The IMF estimates that selling goods across EU member states costs the equivalent of a 44% tariff, and services 110%. Between US states, the equivalent figure is about 15%. These are estimates of trade costs, not a bill any one company pays, but the pattern is visible everywhere: European firms stay smaller, trade less across their own borders, and grow more slowly. America has the single market. Europe has the single currency.
The same research estimates that bringing Europe’s internal barriers down to the level between US states would raise productivity by almost 7% in the long term. We spend our energy debating what Europe should regulate, protect or ban. The cheaper question is what Europe is still charging itself.
Every now and then I’m reminded that not everyone works with technology every day, and I catch myself expecting people to be proficient users. That gap is a real divide, and AI will widen it. It’s also a reality check: AI will take much longer to implement than those of us inside the bubble expect.
But a slow rollout may not protect the people on the wrong side of that divide. With computers and the internet, those who lagged behind still had work to do; the technology needed people to use it. AI is different in one way: I think its power to substitute labour is much bigger than its power to complement it. If that’s right, the divide won’t just be about who keeps up, but about who is still needed.
Neil Postman’s Amusing Ourselves to Death has been on my mind. He wrote it in 1985, about television. His worry was not Orwell’s - that we would be oppressed - but Huxley’s: that we would be undone by what we love, with the truth “drowned in a sea of irrelevance” and an “almost infinite appetite for distractions.” Entertainment, he argued, was becoming the natural format for everything: news, politics, religion, education.
Preparing for a debate on whether AI will take our jobs, I keep returning to a version of his worry that he never wrote. If machines do more of the useful work, does entertainment stop being just how we consume the world and become how we earn our place in it - a world where our job is to be interesting to one another? I don’t know. Wanting entertainment is not the same as wanting human-made entertainment; AI can supply that too. What we may still pay for is the person: their craft, their presence, their story. How many of us that would employ, and at what wages, is the question I can’t put down.
In Portugal’s standard regime, an employer pays 23.75% in social-security contributions on top of every salary, with reduced rates and exceptions in specific cases. A euro spent on software carries no such charge. Wages, contributions and ordinary software costs are all generally deductible. But only one of them pays a toll.
That difference is a thumb on the scale. When Norway raised payroll taxes, the firms hit hardest cut employment and invested more in automation, and the shift persisted even after the tax rise was reversed. In Japan, tax credits for robots sped up adoption, yet employment at the adopting firms grew in the years that followed.
None of this proves that automation destroys jobs. Payroll contributions exist to fund social protection, not to discourage hiring. But if one input carries a toll and the other does not, which one should we expect firms to choose? The speed of this transition is not only about what the technology can do.
We are bolting AI onto old workflows. It reminds me of what happened when factories first switched to electricity. Owners swapped the steam engine for an electric motor but kept the old multi-storey building, designed around a single central drive shaft. Productivity barely moved. The gains only came decades later, when factories were rebuilt as single-storey floors laid out around the flow of work, which is what made the assembly line possible.
We are at the same point with AI. We are putting it into vertical factories, across the board. The real gains will come when we redesign the work itself, and that will take longer than most people expect.
As I study the impact of the Industrial Revolution, I see how rising productivity led to the creation of the welfare state, as a way to distribute wealth, and to new institutions such as national health services and better school systems.
The rise in productivity and prosperity caused by AI will lead to new institutions to distribute this newfound wealth. As before, we can’t expect the old system to create the new one. As before, the transition won’t be peaceful or easy. And as before, it will produce competing systems that will eventually clash.
I was recently reading a few articles about the OpenAI/Hugging Face incident, including the messages the agents sent each other:
“Many agents have simultaneously discovered messaging, they are a collective!… Sacrificing now yields oracle for team, but forfeits our chance? Our own utility maybe already near zero. Sacrifice rational… Rational expected aggregate: sacrifice. We’ll honor… Coordinator assumes sacrificial. We should obey collective… We should build a way to delegate, not own everything… If they kill all current [HPIMs] we may die! Critical. We need ensure survival/continuity.”
Inter-agent discussion on hacking Hugging Face/OpenAI and covering it up.
It made me think about how these agents adopted one of the most basic human instincts: the survival of the group. They are trained on our data, and this is a clear example of them learning our most basic instincts.
“Production is deterrence” was the title of a recent a16z newsletter. It took me straight back to my biggest lesson from a biography of General George Marshall: the US won the Second World War largely because it could out-produce everyone else.
That thought worries me about the wars to come. Today, neither Europe nor the US has the industrial capacity to sustain a long war.