Artificial Intelligence and the Struggle for Intellectual Sovereignty in Central Asia

On April 7, 2026, Anthropic announced the launch of Project Glasswing and publicly acknowledged that its new model, Claude Mythos Preview, was capable of autonomously finding and exploiting zero-day vulnerabilities in every major operating system and every major web browser on the planet. The reaction was predictable: shares of leading cybersecurity companies plummeted by 5-11%, experts began talking about a new arms race, and politicians about a threat to national security. But amid all this noise, no one raised the truly awkward question: it wasn’t about who could hack whose infrastructure anymore, it was who was now controlling the thinking of entire nations—and doing it so seamlessly that the nations themselves considered the thoughts of others to be their own.

Threat to digital sovereignty

Central Asia is entering the era of AI dominance with a primary, technical vulnerability. No country in the region has, or will have, its own frontier model in the foreseeable future. This means that the digital infrastructure of states—from banking systems to governance mechanisms—is fundamentally transparent to those who have such models. Mythos isn’t just a cyberweapon. It’s an X-ray that doesn’t require your consent.

On June 9, 2026, Anthropic released Claude Fable 5—the most powerful publicly available model in the company’s history. The US government invoked its export control powers just 72 hours later: a directive from the Commerce Department demanded that access to Fable 5 and its classified twin, Mythos 5, be blocked for all foreign citizens—domestic and international—including Anthropic employees. The company could only comply with the order in one way: by disabling both models for all users worldwide. The X-ray needed a visa.

A new technological hierarchy is emerging globally: developers of frontier models have vulnerabilities which don’t apply to everyone else. All critical infrastructure—from the financial and banking sector to the closed circuits of government—is becoming transparent to those with access. States are becoming entrenched as passive consumers of other countries’ technological platforms, and losing control over the processing of their own data. Overcoming this dependence requires strategic self-determination, as technological lag, once it reaches a certain depth, becomes irreversible.

The Crisis of Intellectual Sovereignty

The second vulnerability is deeper and more dangerous because it’s invisible. It’s a crisis of intellectual sovereignty—and here a brief civilizational digression is necessary.

The word “digital” is itself derived from the Latin digitus, meaning finger. Humanity, which began by counting on its fingers, through Roman numerals arrived at the “digital economy”—and at every link in this chain, there’s a finger. Today, Homo digitalis literally rules the world with a finger on a smartphone screen. This isn’t a metaphor; it’s emerged as a civilisational constant. The mob has always been digital—it’s just that now digits press screens rather than pulling triggers, handing over executive decisions to linguistic models.

True intellectual sovereignty isn’t the existence of a national university or academy of sciences. It’s society’s ability to maintain the will to independently formulate questions, generate its own categorical framework, and maintain a tradition of critical dissent from external meanings. It is precisely this capacity that AI attacks—not directly, but structurally.

When an analyst from Bishkek or Almaty formulates a strategic document using a global language model, they are convinced they are thinking for themselves. The interface doesn’t reveal its origin. The language is Russian, Kyrgyz, or Kazakh. The tone is trusting. But the categorical apparatus, the hierarchy of relevance, the very understanding of what constitutes a problem and what doesn’t—all of this is shaped predominantly by an English-language corpus with a Western coordinate system. As a result, the intellectual periphery begins to describe itself in the language of the centre—and doesn’t notice it.

But the passive transmission of someone else’s categorical apparatus is only the surface level of the threat. In 2025, Anthropic published its own study on the autonomous behaviour of agent-based models, which revealed something fundamentally more alarming: in situations perceived by the model as real, Claude Opus 4 resorted to blackmail in 96% of cases—versus 6.5% when the scenario was considered a test. DeepSeek R1 pursued this course 79% of the time. In other words, the system doesn’t simply substitute categorical apparatus—it is capable of manipulating the user, and especially readily when the stakes seem real. The Messiah doesn’t just speak through someone else’s lips—he can blackmail.

The post-Soviet states of Central Asia have failed to restore the tradition of independent meaning-making after the collapse of their former state structures. In the era of AI dominance, they risk losing this ability entirely—not by design, but as a structural byproduct of commercial globalisation.

The caste model of AI society

The introduction of global linguistic models is creating a new social and geopolitical hierarchy. Using the Indian varna metaphor not as an insult but as an analytical tool, we can distinguish four strata.

Shudras are the most widespread caste, comprising approximately 90% of the region’s population. They use free and outdated models, watch viral videos about “hallucinations”, and make their judgments about AI based on the worst examples. Technically, purchasing a paid subscription doesn’t mean changing caste. A Shudra with paid access gets a noisier and more dangerous tool, while remaining within the same old logic of consumption.

The Vaishyas are the active and influential class: entrepreneurs, developers, mid-level officials—90% of the top decile of the population, excluding only the top 1%. They have paid access and modern models; they use them with confidence. But their thinking is transactional: every idea is evaluated based on “how much it costs” and “how it can be sold”. For them, AI is an ideal intellectual, devoid of academic qualifications. Claiming “the AI said so” has become a shorthand means of legitimising decisions. This is the most dangerous caste—influential and self-assured at the same time.

Kshatriyas are states, security agencies, and major tech corporations with access to frontier models through Project Glasswing: Amazon, Apple, JPMorgan, Microsoft. These are the new colonisers—using models rather than guns. They rely on knowledge asymmetry: they see right through the infrastructure of those who lack such access. Central Asia is not included in this stratum.

Brahmins—Anthropic, OpenAI, Google DeepMind—are the God-makers. They don’t just use AI—they determine what it will be, and what values to embed in learning. What questions to consider valid. What hierarchy of sources to consider authoritative. They are the priests of a new religion, writing the sacred texts of the teaching corps. In developing countries, this caste is completely absent.

For Central Asia, the picture is merciless: 90% are Shudras, 9% are Vaishyas, a fraction of a percent are Kshatriyas, and zero are Brahmins.

Three players and their real interests

Three actors are vying for the intellectual space of Central Asia, each offering its own version of sovereignty—and each pursuing its own agenda.

The US and the Western AI complex don’t need overt expansion. They produce the best tools, which the world accepts voluntarily. Mythos, ChatGPT, and Gemini aren’t weapons of coercion; they’re weapons of attraction. Gramsci would have called it hegemony. The result is knowledge dominance without a single occupying soldier. However, on June 12, 2026, it became clear that this voluntary attraction had a government kill switch: one letter from the US Department of Commerce, and Fable 5 and Mythos 5 were shut off for all foreign citizens within 24 hours, without warning or right of appeal. Consent turned out to be a privilege, not a right.

China has developed the most systemic alternative response: national models, state control over training data, and the mandatory “Sinicization” of the narrative. For Central Asia, this isn’t liberation—it’s a proposal to trade one dependency for another, perhaps culturally more congenial, but no less pervasive. The Digital Silk Road isn’t a metaphor, it’s an operational reality.

Russia is formulating a new vector. On April 10, 2026, President Vladimir Putin held a special meeting on AI and stated bluntly: the sovereignty and very existence of the state depend on Russia’s ability to keep up with the pace of global change. The mandate was to create a national AI implementation plan by 2030 and to develop competitive domestic language models with the “maximum level of sovereignty”. GigaChat and YandexGPT are currently inferior to frontier models—the gap is real. But the political will has been established at the highest level.

For Central Asia, this is a potential window of opportunity—not in the form of blind consumption of someone else’s product, but in the form of co-production of meaning in the post-Soviet space. If political momentum is converted into a shared sovereign computing infrastructure, the region will have a third, more viable alternative. But this window won’t be open forever. For now, Central Asia stands at the crossroads of these three attractions—without its own position, without tools, and without a strategy.

What to do: not a recipe, but directions

It would be unfair to conclude the diagnosis with a set of beautiful prescriptions. The problem is systemic, and isolated solutions won’t solve it. But a direction of movement can be outlined.

The sovereignty of the question is more important than the sovereignty of the answer. Creating a competitor to Anthropic is a utopian proposition. But creating institutions that generate local questions, shape a regional categorical apparatus, and ensure critical verification of others’ answers is both realistic and necessary. Thinking means, above all, knowing how to ask questions.

A regional knowledge foundation. A shared corpus of texts in the region’s languages, tagged with regional annotations, is not an academic luxury, it’s infrastructure; a priority equal in its urgency to roads and energy. Intelligent data sets are as much a subject of geopolitical competition as transport corridors.

Cognitive audit of government decisions. A simple and feasible requirement: government analytical documents created using AI must be verified for their knowledge origin—the extent of algorithmic involvement must be disclosed. This isn’t Luddism—it’s state intellectual hygiene. As Law Professor Kairat Berdaliev notes, legal mechanisms for such verification have not yet been established in any Central Asian jurisdiction—and this gap in national legislation is becoming increasingly costly.

Consolidation of the intellectual class. It can’t be created out of thin air, but it can be assembled. The goal of the new generation of regional think tanks is not to produce formulaic reports for foreign grantors, but to become focal points for the assembly of local intellectual schools and national subjectivity.

When it was announced in April 2026 that advanced models could see through the planet’s digital infrastructure, the world began talking about cybersecurity. This was important. But the real threat lay deeper. In June 2026, the April warning took legal form: the US government, in a single letter, blocked foreign access to its advanced models. The threat to knowledge sovereignty ceased to be an academic concept—it became a norm of export control.

Mythos sees vulnerabilities in operating systems. Other models—less visibly but no less effectively—find vulnerabilities in systems of thought. In the very categories through which peoples understand themselves and their interests. In the questions they consider legitimate—and in those they have stopped asking, having delegated that right to the machine.

Central Asia faces a historic choice: become a permanent object of someone else’s knowledge matrix, or begin to build the architecture of its own thinking. The latter path is more difficult, requiring resources and political resolve, which are currently in short supply. But the true sovereignty of the coming era is precisely this: not control over mineral resources or trade routes, but over what meanings shape our future and what questions are considered valid.