"With AI, waiting is no longer an option"

In the past, companies were run by those who mastered their technology. With the advent of modern corporate structures, the two began to drift apart. Technical expertise shifted to the executive boards and specialist departments. Capital, legal affairs and networking moved to the supervisory boards. For a long time, this was manageable: a steam turbine from yesterday was a steam turbine of tomorrow. Purchasing one was a matter of calculation, not a gamble.

With artificial intelligence, this division of labour no longer works. Anyone who does not understand what AI systems can and cannot do cannot adequately assess the risks, dependencies and strategic significance. Anthropic’s decision to withdraw high-performance models from certain markets has shown that: Whoever controls the systems controls access. And Germany controls neither one nor the other. 

Access to state-of-the-art AI is not a technical issue. It is a question of power. There is one problem that is conspicuously rarely raised in the German AI debate: the organisational problem. Many companies and institutions are not organisationally equipped to keep pace with the speed of technological change. Decision‐making structures that have been optimised for stable markets act as a brake in highly dynamic environments. 

This is particularly evident in supervisory boards. These are predominantly made up of lawyers, finance experts, politicians and experienced managers. All incredibly clever people with impressive CVs. However, expertise needed to make informed judgements on AI systems, models or technical governance risks is rarely found among them. And those who do not understand technology cannot manage it. 

We are already familiar with this pattern from the cloud and blockchain: first ignorance, then hype, followed by scepticism and finally belated adoption. With AI, this is more difficult. AI is already a basic infrastructure, just like gas, electricity or the internet. Anyone who does not build, master and control this infrastructure will become dependent. 

Germany does not have a technology problem, but rather a distribution problem, an institutional problem and a problem of understanding. There is a difference between regulation that supports a trustworthy AI ecosystem and regulation as a substitute for a strategy. Too often, it is the latter. A sensible approach to AI requires forward-looking and, above all, sociotechnical consideration: AI encompasses not just models, data, and computing power, but also adoption, integration, user acceptance, compliance architectures, and governance instruments.

The development of digital systems moves like a pendulum between centralization and decentralization. From mainframes to PCs to the cloud and now to highly concentrated AI platforms, control and power repeatedly shift between a few actors and broader structures.

We are currently experiencing a phase of extreme centralization—centralization of AI systems and, with it, of control over AI development and use. What this means concretely: we are subject to external control over who gets access, under what conditions, under which laws, and with what possibilities. Anthropic's decision is not an isolated case. The pendulum will swing back; it always does. The question is whether this shift is actively shaped or only happens as a reaction to crises. This requires two levels:

• The technical level: Who operates the systems, who trains the models, on what data, in which data centers, and under what law.
• The societal level: What decisions are AI systems allowed to make, who is liable, and what values are anchored?

We write traffic rules for cars we don't build, on roads we don't own. That is neither system governance nor usage governance. And it is certainly not sovereignty.

Cohere's acquisition of Aleph Alpha exemplifies a German problem: Germany produces talent and ideas, but control migrates to North America. Yet there are also positive examples. Schwarz Digits, for instance, is pursuing a different path with Stackit. With an investment of eleven billion euros in a data center in Lübbenau, an infrastructure is being built that is subject exclusively to European law.

Stackit is not yet on a par with Amazon Web Services (AWS) technically. That carries real weight. But it weighs even more heavily that without its own providers, Germany would have no choice at all. Because even locally stored data is subject to certain legal access rights of their home countries when held by US providers. Sovereignty means freedom of choice. Even with Stackit, more comparable German alternatives are still needed.

Trust is created through system design, not through declarations of intent. Here lies the deepest misunderstanding in the German AI debate. Trust is treated as a moral category. But trust between people, institutions, and AI systems does not arise from stated intentions—it arises from system design and governance that justify that trust.

In areas such as medicine, justice, or credit lending, there are structural tensions: greater accuracy often means less fairness. Better data protection frequently costs predictive performance. Higher explainability can reduce precision. These tensions cannot be eliminated. They must be addressed with social considerations in mind, not just technical possibilities. This requires decision-makers who understand both the technology and its societal consequences. Whoever fails to establish this competence on boards, supervisory boards, and in public authorities loses the power to shape outcomes.

Germany can show that trustworthy AI and competitiveness are not a contradiction, but rather strengthen one another. And if it works for Germany, it can work for all of Europe too. What needs to happen now:

• First: Supervisory boards and executive boards need technical competence as a core requirement. This does not mean that every board member must be able to build an AI agent. It means that at least some members of decision-making bodies must be able to assess system architecture, models, data, and technical and technological risks.
• Second: Institutions need the capacity to learn. Legislation on five-year cycles is too slow for AI. This calls for new governance formats: regulatory bodies with technical expertise, adaptive rule frameworks, and bodies that anticipate rather than merely react.
• Third: A dual strategy of technical and societal governance is needed. On one hand, investment in European computing capacity, AI models, and technological independence; on the other, governance structures that actively accompany socially desirable technological development.

When it comes to AI, waiting is no longer an option. The question is not whether. The question is: under what conditions, under which laws, with what values, controlled by whom? Digital sovereignty arises from political will and becomes reality through design decisions. The pendulum is swinging. Whoever understands its dynamics has a clear advantage. Whoever ignores it will be struck by it.

(This article was translated with AI) 

Link to the Article: 
https://table.media/research/standpunkt/bei-ki-ist-warten-keine-option-mehr?utm_source=share&utm_medium=social&utm_campaign=article_share

PDF: “With AI, waiting is no longer an option”