When large language models enter the conversation, they are most often understood as tools that produce text from a handful of prompts. That perception is not wrong, yet it is dramatically incomplete. The year 2022 was the moment these systems entered our daily lives; from the vantage point of 2026, we are witnessing not a plateau but a continuing ascent. The big technology companies release model after model, some popular, some affordable, all consequential. Yet the more decisive development is unfolding elsewhere. States themselves have begun to build their own models, and in the technopolar reality that I have examined in detail in my previous columns, this is not a matter of technological prestige. It is a matter of digital autonomy, and behind digital autonomy stands a deeper question that I have conceptualized as epistemic dependency.

Why does an indigenous model matter so much? The answer begins with the nature of "dependency” itself. Energy dependency is visible: Pipelines can be mapped, contracts can be renegotiated, suppliers can be diversified. Technology dependency is measurable: chips, licenses and patents leave a paper trail. Epistemic dependency, in contrast, is largely invisible, and precisely for that reason it is far more insidious. When a society’s knowledge environment is mediated by foundation models trained elsewhere, on corpora it did not curate and under values it did not choose, that society lives in the condition we have described in my previous writings as epistemic dependency. The categories through which it interprets the world, the sources a model treats as authoritative, the questions it declines to answer and the framings it quietly normalizes are all decided at the point of design. A state that analyzes its own data with someone else’s algorithm has, in practice, transferred the sovereignty of that data. Over time, this silent architecture shapes foreign policy preferences, national security assessments and strategic decision-making without ever announcing itself.

The threat is graver than it first appears, because the deepest information harms of the generative era do not occur at the moment of output. They occur at the moment of training. If a foundation model ingests falsehoods, distortions, selective archives or fabricated historical claims, it does not merely repeat error; it industrializes it. Defects at the foundation do not remain at the foundation. They travel into every downstream application that treats the model as authoritative, from search to education to journalism, and a "lie” that enters the training corpus can reemerge as machine-readable orthodoxy. Amplification and generation may be legally distinct categories, but they converge in effect: both determine the epistemic conditions under which entire publics reason.

This is also why an indigenous large language model is never merely a linguistic achievement. A model trained on a nation’s own corpus does not simply translate a mother tongue into vectors; it translates a nation’s culture, values, traditions, history, collective memory, future expectations and even its feelings. Proverbs, idioms and the emotional grammar of a society are either encoded faithfully or flattened into someone else’s approximation. In the technopolar world, a country’s true and annotated story can only be told through a national, indigenous model. Everything else is a story told about her, not by her.

That this argument has become state practice is now beyond dispute. In the Balkans, Bulgaria’s BgGPT was developed by the INSAIT institute with government backing; its founder has placed a $100 million bet, precisely so that data would be stored and processed at home. In Eastern Europe, MamayLM, a Ukrainian model built by researchers at INSAIT and ETH Zurich, was designed to think in Ukrainian at a fraction of the cost of frontier systems. In the Gulf, the United Arab Emirates’ Falcon series showed that a national institute could produce openly documented models competitive with the largest laboratories, while France’s Mistral became Europe’s flagship answer to Silicon Valley. Different regions, one shared diagnosis: whoever outsources the model outsources the meaning.

Türkiye has read this diagnosis correctly, and her answer now has a name: Bilge. Announced by President Recep Tayyip Erdoğan at the unveiling of Türkiye’s Artificial Intelligence Vision and Action Plan, Bilge is a family of large language models developed by TÜBITAK BILGEM under the auspices of the Ministry of Industry and Technology. The president described the model as an important indicator of the distance Türkiye has covered on this road. The Bilge family spans a wide parameter scale, from one billion to 122 billion, with versions of 1.9 billion and 27 billion parameters already built and a 122 billion parameter model still under development for complex reasoning and multistep problem solving. The model has not yet been opened to general public use; according to Turkish media reports, its debut in state service will include a role as the president’s personal artificial intelligence assistant. The same reports, citing official statements, note that Bilge was engineered end-to-end by TÜBITAK BILGEM with domestic resources and that it has led the general translation category against prominent international models. These figures are declared milestones, not finished monuments.

Read through the lens of epistemic dependency, Bilge’s significance is categorical rather than incremental. A model that thinks in Turkish, trained on the distinctive structure of the language and the cultural accumulation it carries, is an instrument through which Turkish society can interrogate the world on her own terms, anchoring the concept Turkish tradition captures in a single word, "hakikat,” inside the computational layer itself. Nor does the vision end at Türkiye’s borders. As noted in my previous writings, the planned Large Turkic Language Model, to be developed with the partnership of the Organization of Turkic States and encompassing the Oghuz, Kipchak and Karluk branches, aims to build a common epistemic ground for the entire Turkic world in the digital age. Sovereign models, in this vision, are not walls; they are bridges built on one’s own foundations.

There is, finally, the classical public diplomacy dimension. Public diplomacy has always been the art of telling a nation’s story to foreign publics, from radio broadcasts to cultural institutes. In the algorithmic age, however, the first narrator a curious foreign student, journalist or diplomat consults is no longer a broadcaster but a model. If that model has learned Türkiye only through external archives and secondhand framings, then her story is being told, billions of times a day, by narrators she never appointed. An indigenous model reverses this asymmetry. It is public diplomacy compiled into weights and parameters, the point where storytelling becomes what I have conceptualized as cognitive diplomacy: the deliberate cultivation of presence, practice and resilience in the cognitive infrastructures where international relations now unfold.

None of this implies isolation. Türkiye will continue to engage global models, markets and governance platforms, where she is already emerging as a norm entrepreneur for human-centered artificial intelligence. But engagement from a position of epistemic sovereignty and engagement from a position of epistemic dependency are entirely different postures. Bilge signals that she has chosen the former. The candle of borrowed narratives, as the Turkish proverb reminds us, burns only until the evening prayer. Nations that intend to be understood truthfully in the technopolar century must light their own. With Bilge, Türkiye has struck the match.