Apply AI: Europe's AI strategy has forgotten European values
A recent announcement of the European Apply AI summit made me reflect (or rant) again on the Commission's Apply AI Strategy, which soon will turn one year old. I reread the strategy, and one passage got to me again: AI should be considered as a potential solution "whenever organisations make strategic or policy decisions", under an "AI first policy".
As I've repeatedly said, this is the wrong starting point. Question Zero comes first: what are we trying to achieve, for whom, and is AI the right means at all? Under what conditions, if any, should a system be adopted? Building and regulation come after, not before. "AI first" asks how fast the technology can be deployed. The strategy does mention weighing benefits and risks, but a default in favour of AI places the burden of proof on those who doubt. A legal order based on European values should place it on the deployer.
"AI-first" inverts good policy practice. Sound decision-making starts from the problem, the affected people and the available options, and only then asks which tools fit. The evidence is substantial. The Dutch childcare benefits scandal and Australia's Robodebt show what follows when automation precedes understanding. Starting from the technology is the classic law of the instrument. Productivity is not the measure of AI's benefit. Faster is not better, and a system can gain speed while losing fairness, accountability and trust. Progress should be measured by wellbeing and rights, not output. Sometimes the right answer to Question Zero is not to use AI at all, and a serious strategy should be able to say so.
The aims are equally telling. Competitiveness, adoption and technological sovereignty dominate. These matter, but they are not what made the European approach distinctive. "Trustworthy, human-centric" appears in the summit's tagline, yet the draft agenda gives the AI Act, workforce and safety one parallel session each, alongside energy, manufacturing and defence.
"Buy European" is welcome in principle, and so is the preference for open source. But made in Europe is not sovereignty. Sovereignty means control: over the data, the compute, the model and the ability to inspect, adapt and switch. A system hosted in Europe, built on foreign infrastructure and opaque weights, delivers dependency with a European label. Open source enables inspection, adaptation, independent audit and the freedom to switch provider, which is what sovereignty actually means. But not all "open" AI is truly open: research on open-washing shows weights released without training data, documentation or the right to modify. Procurement should therefore demand genuine openness and test purpose, control, accountability and environmental cost, not only origin.
Europe does not need to follow a direction set by others. The future of AI is not something that happens to us, it is something we choose. Frontier AI is more than scaling data-driven models, and it is a choice to equate the two (a choice that conveniently suits US-based big tech). Europe can set a different direction: AI that is purposeful, governed and built on diverse approaches, adopted where it serves people and left out where it does not. That is the debate I would gladly see on the main stage in November.
