The Second Front of AI Sovereignty

The first sovereignty panic was easy to understand. If a foreign government can squeeze a model provider, then advanced AI starts to look less like software and more like strategic infrastructure. But that is only the first front. Even a model that remains perfectly available can still speak with someone else's instincts.

That first front became concrete in June 2026. On 12 June 2026, Anthropic said a US export-control directive had forced it to suspend access to Fable 5 and Mythos 5 for foreign nationals. On 26 June 2026, OpenAI said GPT-5.6 would begin with a limited preview for a small group of trusted partners at the US government's request before broader rollout. Those two moments did not prove that every organization now needs its own national model stack. They did something more useful: they made the access axis of AI sovereignty tangible.

The earlier article AI Sovereignty Is Not a Datacenter asked who holds the off-switch. This second question is harder because it survives even when the switch disappears. Suppose Europe trains or hosts a model nobody in Washington can remotely withdraw. Suppose access continuity is solved. The deeper dependency can still remain intact, because the system may continue to carry imported assumptions about authority, religion, social norms, fairness, acceptable dissent, and what counts as a reasonable answer.

That is the second front of AI sovereignty. Not who can interrupt the system, but whose worldview stays switched on all the time.

Values are not a metaphor

One reason this argument is worth making now is that the research base has matured. A few years ago, "model values" could still sound vague, almost literary. Today the safer claim is the opposite: values are visible enough to be measured, compared, and manipulated.

Anthropic's Values in the Wild paper identified 3,307 AI values across real-world production interactions. The significance is not the number itself. It is the fact that a deployed system repeatedly expresses normative priorities in context, not just technical competence. ValueCompass pushes the point further by offering a framework to compare human and model value alignment across domains such as healthcare, education, writing, and the public sector. The picture that emerges is not of a neutral machine occasionally drifting into politics. It is of a product that carries practical, moral, and epistemic defaults into ordinary use.

That matters because the values layer enters from at least two directions.

The first is the training distribution. Frontier models are trained on web-scale corpora in which English and Anglophone internet culture remain structurally dominant. The strongest research does not justify a cartoon claim that all Western models simply speak with one tidy liberal-secular ideology. It does support a narrower and more serious point: these systems inherit background assumptions from the data mix that shaped them, and those assumptions often make dominant Western styles and habits feel more natural than local ones.

The second is post-training steering. Instruction tuning, reinforcement learning from human feedback, moderation layers, safety rules, refusal policies, domain tuning: all of these shape not only whether a model answers, but how it frames an answer, what it avoids, and which trade-offs it presents as sensible. Studies on ideological fine-tuning and political moderation show that these layers can move outputs in measurable ways. In other words, the worldview is not just sediment from the corpus. It is also policy.

That is why value sovereignty is separate from access sovereignty. A model can remain available and still influence everyday thinking through imported norms.

Influence rarely arrives dressed as propaganda

The cliché version of the debate imagines ideology entering AI the way censorship enters a newspaper: loudly, visibly, with obvious red lines. That does happen in some systems and on some topics. But the larger influence channel is usually much duller, which is exactly why it is dangerous to underestimate.

A model does not need to lecture users about civilization, morality, or state doctrine to shape a society's defaults. It only needs to become the standard co-writer, explainer, summarizer, classifier, and adviser. Once that happens, small suggestions accumulate. A phrase becomes more natural. A risk is framed one way rather than another. One moral conflict is treated as obvious, another as too sensitive, a third as already settled. The cumulative effect can be cultural without ever becoming theatrical.

That is also why the CHI 2025 paper by Dhruv Agarwal, Mor Naaman, and Aditya Vashistha, AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances, is useful here. Their controlled experiment with Indian and American participants shows a mechanism of influence that is mundane rather than melodramatic. Nobody needs to ban local culture. The model merely helps millions of users write, and in helping them it quietly bends expression toward the defaults embedded in the model.

This is the point Europe cannot evade by buying more GPUs. Compute sovereignty may reduce the risk that someone else can deny access. It does not, by itself, answer who authored the norms inside the system Europeans end up renting every day.

The phrase problem inside "European values"

This brings us to the most interesting point. Europe already has a vocabulary for this debate. OpenEuroLLM invokes European values, transparency, openness, multilingualism, and diversity. The European Commission's AI framing ties industrial capacity to trust, safety, and fundamental rights. The older Ethics Guidelines for Trustworthy AI provide a more concrete grammar still: human agency, oversight, transparency, fairness, accountability, societal well-being, privacy, and robustness.

None of that is empty. It would be wrong to sneer at it as pure branding. Europe does possess a real rights-and-governance language for AI.

But it is also not yet a full behavioral doctrine for general-purpose models. The official material is much clearer on compliance, auditability, transparency, and legal principle than on the everyday normative collisions that make values real. What should a model do when pluralism and safety conflict? When anti-discrimination and religious doctrine collide? When a user asks for arguments against immigration, feminism, same-sex marriage, or EU integration? Should a public-sector model mirror prevailing law, median public opinion, charter-level rights language, or some curated mixture of all three?

This is where the parallel with the energy debate in Wie durft te kiezen? becomes useful. "Energy mix" sounded concrete while often postponing the real argument over allocation: which technologies, in what proportions, under what time horizon, at whose cost? "European values" can play the same role in AI. It sounds substantial. It signals seriousness. It may still conceal a refusal to specify the actual precedence rules.

Which values? In which domains? With what trade-offs? Under whose authority?

Until those questions are answered more operationally, "European values" remains directionally meaningful but strategically incomplete.

The hard counterargument deserves a direct answer

The reason for this line of thinking is also its strongest objection. Europe wants its own value layer partly to become less dependent on American or Chinese defaults. But once it starts embedding European values into AI systems, the hard question becomes unavoidable: how is that different from what Europe criticizes elsewhere?

The lazy answer is civilizational flattery: Europe has values, China has censorship. That answer is too crude to survive contact with the literature. Comparative work on moderation shows that all major ecosystems steer models. Western, Chinese, and Russian systems all draw lines. They all nudge, omit, refuse, and rank sensitivities differently. No serious observer should still pretend that one bloc has politics while another has neutrality.

The real distinction, if Europe wants one, has to be procedural.

A publicly governed values layer is not the same thing as opaque state-protective censorship. The relevant questions are whether the rules are stated, whether they can be contested, whether independent auditors can inspect them, whether the rationale is publicly justified, whether the system can be revised, and whether the model remains free to criticize the governing order itself. That is a meaningful line. It is not a perfect one, but it is defensible.

Put differently: the difference is not that one system has values and another does not. The difference is whether the value layer is politically owned in the open or politically imposed behind the curtain.

That still leaves Europe with an uncomfortable problem of its own. The continent has regulation, institutional vocabulary, funding initiatives, and a strong legal imagination. What it does not yet seem to have is a settled public mechanism for translating high-level rights language into model behavior in morally contested cases. If the answer is "the companies decide," value sovereignty collapses back into vendor ideology. If the answer is "Brussels decides," critics can reasonably worry about technocratic worldview management. If every member state decides for itself, the result may be fragmentation under a shared slogan.

That gap is precisely why the issue matters. Europe may end up talking fluently about value sovereignty while outsourcing the actual value layer to whoever fine-tunes the model.

The second sovereignty question is about authorship

This is the deeper reason this argument belongs next to AI Sovereignty Is Not a Datacenter, rather than underneath it. That earlier piece argued that AI sovereignty is not a datacenter. It is a layered choice about acceptable dependence. That remains true here. But the second sovereignty question forces a more awkward recognition: even when you solve infrastructure dependence, you may still be renting someone else's moral defaults.

For a hospital, that could shape how risk is explained to patients. For a school system, how history or sexuality is framed. For a ministry, how trade-offs are summarized. For a newsroom, which formulations feel neutral and which feel loaded. None of that requires spectacular censorship. It only requires daily use at scale.

So the sharper version of the European AI debate is no longer just whether the continent can train, host, or procure its own systems. It is whether Europeans are willing to specify, govern, and contest the worldview embedded in the systems they increasingly rely on.

The first sovereignty debate asked who can interrupt the machine.

The second asks who quietly governs it while it keeps running.

Sources

Access sovereignty / June 2026 trigger

• Anthropic statement on Fable 5 and Mythos 5 access suspension — export-control directive and foreign-national access restriction
• OpenAI GPT-5.6 Preview System Card — government-requested limited preview language
• TechCrunch on the GPT-5.6 rollout restriction — reporting context around the OpenAI preview decision

Primary research on value measurement and alignment

• Values in the Wild — large-scale extraction of AI values from production interactions
• ValueCompass — framework for measuring human and model value alignment across domains
• Whose Opinions Do Language Models Reflect? — demographic misalignment between model outputs and human opinion groups
• PoliTune / data selection and fine-tuning effects — measurable shifts in political and economic bias
• Ideological manipulation via fine-tuning — how small slanted instruction sets can move output behavior

Cultural skew and downstream influence

• How Well Do LLMs Represent Values Across Cultures? — cross-cultural value representation
• Fluent but Foreign: Even Regional LLMs Lack Cultural Alignment — regional fluency versus cultural fit
• AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances — everyday assistance as a vector of cultural standardization

Moderation, censorship, and geopolitical steering

• What Large Language Models Do Not Talk About — comparative moderation across geopolitical ecosystems
• R1dacted: Investigating Local Censorship in DeepSeek's R1 Language Model — state-sensitive suppression in a Chinese-origin model
• Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation — evidence that models can know and still suppress

European initiatives and official value language

• European approach to AI — EU framing around excellence, trust, and rights
• Ethics Guidelines for Trustworthy AI — the clearest formal EU value scaffold
• OpenEuroLLM official site — stated commitments around openness, multilingualism, and alignment with European values
• OpenEuroLLM launch press release — official positioning of the initiative
• STEP / OpenEuroLLM project page — EU project framing and objectives