Germany has found a solution to artificial intelligence. Before the machine becomes cleverer than us, we shall make the ownership structure more complicated. If humanity must face extinction, it should at least know which ministry is responsible.
In his September 18 ZEIT column, Marcel Fratzscher proposes majority state ownership of all AI models and their infrastructure. Private partners and competition may remain, but ultimate control belongs to the state. The apocalypse, apparently, becomes manageable once it has a supervisory board. ZEIT column.
His concerns deserve consideration. Concentrated corporate power, political manipulation and dangerous capabilities are serious problems. But the argument skips the essential step: showing why changing the shareholder register would solve them. A government stake does not make a model truthful, secure or controllable. Public ownership is an institutional arrangement, not an alignment technique.
The scope is extraordinary, too. A small classifier and a frontier system capable of planning complex operations are both AI models. Treating ownership as the universal safety switch flattens precisely the distinctions that serious risk management requires. The proposal is simultaneously sweeping about property and vague about implementation. That is quite an achievement for something presented as a concrete solution.
The nuclear analogy makes this especially awkward. The nuclear arsenals that threatened humanity were built and controlled by states. Whatever else that history demonstrates, it cannot establish that state ownership removes catastrophic incentives. Governments also compete, conceal failures and invoke national security. The mushroom cloud does not become reassuring because the paperwork carries an official crest.
Then there is the evidence. Fratzscher cites more than 1,000 suspicious drone overflights and 321 suspected sabotage cases in Germany in 2025. His linked source reports those incidents; it does not establish that AI caused them. A suspicious drone is not a peer-reviewed demonstration of a language model at the controls. Hybrid warfare is real. Attaching “AI” to its statistics does not complete the causal argument. The cited DGAP page.
Similarly, an employee’s estimate of an extinction risk above ten per cent is a judgement under uncertainty, not a measured accident rate. It warrants investigation. It does not tell us whether the optimal government shareholding is zero, 51 or 100 per cent. There is no equation in which existential dread automatically converts into public equity.
The democratic argument has an even stranger twist. Fratzscher fears political influence and governments such as Trump’s, then proposes putting ultimate control of AI in government hands. Which government’s hands? A durable safeguard must work after an election won by people its designer dislikes. “Trust the state” needs a rather substantial appendix specifying who controls it, who can challenge it and what happens when it turns hostile.
Europe, meanwhile, is hardly an untouched regulatory wilderness. The AI Act already imposes obligations on general-purpose model providers, with additional requirements for systemic risks. It also contains partial exemptions for qualifying open models. That distinction matters: sensible criticism should target disproportionate burdens, not pretend every chatbot needs a Brussels operating licence. European Commission guidance.
But Germany and the EU have provided ample material for satire. Germany reached its Fourth Bureaucracy Relief Act: even cutting red tape has become a franchise with sequels. The Commission itself has targeted administrative-burden reductions of 25 per cent overall and 35 per cent for smaller businesses. When the regulator advertises how much regulation needs simplifying, requesting another layer of control deserves at least a raised eyebrow. German legislation, Commission targets.
The economic concern is straightforward. Fixed compliance costs can favour incumbents. A giant can hire a department; a small developer has to become one. Rules intended to tame dominant firms can end up protecting them from competitors. Before making this machinery more elaborate, an economist might ask who can afford to operate it.
Fortunately, the choice extends beyond billionaire-controlled clouds and ministerial supervision. Downloadable models let universities, businesses, cooperatives and individuals operate useful AI themselves. They cannot all afford frontier training runs. They do not need to train every model they use, any more than every driver needs a car factory.
Here comes the pleasing irony: Chinese developers supplied practical examples back in 2025. Qwen3 released model weights under Apache 2.0 and documents local deployment. DeepSeek-R1’s code and weights carry an MIT licence, with self-hosting instructions and smaller distilled models available under their applicable licences. These are concrete alternatives to depending exclusively on somebody else’s service. Qwen documentation, DeepSeek repository.
A properly configured offline installation need not send prompts to the original provider. That does not erase learned biases, make every model trustworthy or put the largest systems on an ordinary laptop. Hardware, security and evaluation still matter. What changes is the distribution of operational control: more people can run, examine, adapt and replace the tools they use.
That is a practical form of technological autonomy. Its usefulness depends on what users can control, rather than the nationality printed on the download page.
Fund public research and shared computing. Enforce competition. Require proportionate safety measures for dangerous capabilities and deployments. Preserve room for independent developers and local operation. Public investment can widen access without requiring the state to own every model.
If concentrated power is the problem, distribute the power. There is something deliciously awkward about a German proposal to democratise intelligence through state ownership arriving while Chinese developers offer models people can run themselves.
Perhaps the revolution will be downloaded before the competent authority has finished designing its application form.
No comments yet