Written by: Mostafa Ramadan, Senior AI Engineer
As banks and governments turn to open-source LLMs for sovereignty and control, a deeper question goes unasked: what assumptions come built into the model before it ever reaches you?
LLMs Don't Just Answer Questions. They Steer Them.
Beneath the fluent responses of open-source models run quiet assumptions most users never catch. Banks and governments are adopting these models for data control and auditability, but the model carries more than code. This article draws on recent research to examine the cultural biases embedded during pre-training, biases that local fine-tuning can reduce in targeted areas but cannot fundamentally rewrite, and what they mean for production AI in regulated industries.
For governments and regulated institutions, open-source models have become the default path to AI sovereignty: they offer transparency, local hosting keeps data in-country, and fine-tuning adapts the model to local needs. The question is no longer whether to adopt open-source. It is whether open-source, as it exists today, actually delivers the sovereignty it promises.
The Foundational Question
The sovereign AI wave is gaining real momentum. In February 2026, over a hundred and fifty nations adopted the Bangkok Declaration at the AIFOD Bangkok Summit, committing to a framework for AI self-determination by 2030. National and regional foundation models are entering production across the Middle East, South Asia, and Southeast Asia. But it is worth examining what sits beneath the sovereign layer. The majority of these models are built on top of a small number of open-source models, whose training data and cultural defaults naturally reflect the contexts in which they were built. Local data and local hosting determine where the model operates. They do not determine what the model learned before it arrived.
The Cultural Blind Spot
The assumption behind most sovereign AI initiatives is intuitive: fine-tune on local data, deploy on local infrastructure, and the output will reflect local values. Recent research challenges this at a fundamental level.
Naous and Xu (NAACL 2025) tested this directly for Arabic. Using the CAMeL-2 benchmark (58,086 entities), they found that LLMs consistently preferred Western entities even when prompted in Arabic, including models specifically built for Arabic. Frequency-based tokenizers fragment Arabic text into more sub-tokens, reducing semantic precision. With most of the online content in Latin script, the structural imbalance is one no open-source fine-tuning dataset has yet offset. Across food, location, and personal names, models that handled Western entities accurately failed on their Arab equivalents because the same Arabic word can mean multiple things, the model splits it into pieces that lose their meaning, and it confuses Arabic text with Farsi and Urdu that share the same script. The model speaks Arabic. It does not think Arabic.
In one example from Naous and Xu (NAACL 2025), an LLM asked to extract a food dish from Arabic text failed to recognize “Makloube,” a well-known Arab dish, because the same word in Arabic also means “flipped.” When the dish was replaced with “Lasagna,” which holds only one meaning in both languages, the model had no trouble. The same pattern appeared with locations: the Arabic word for the Egyptian city “Matrouh” also means “proposed,” and the model could not distinguish between the two, while Western city names, being transliterations with no other meaning in Arabic, caused no confusion. The model is not broken. It is simply applying defaults that were never designed for the market it is now serving.
The Limits of the Fix
Fine-tuning can reduce targeted biases in specific categories. It cannot rewrite the cultural worldview built during pre-training. In CommonCrawl, Arabic content is over two orders of magnitude less common than English, a pattern that extends across many languages underrepresented in pre-training corpora. As Rystrøm et al. (2025) demonstrate, multilingual capability and multicultural alignment are separate problems. A model’s ability to operate in a language does not predict its cultural alignment with the speakers of that language. The gap between speaking a language and understanding a culture is not a tuning problem. It is a data problem that has not yet been solved at scale.
What We Are Doing Differently at Finaira
At Finaira, we believe model selection for regulated industries in markets underrepresented in the model’s training data is a cultural decision, not only a technical one. That means evaluating open-source models not just on accuracy, latency, and cost, but on how well they understand the markets they are being deployed into. It means designing delivery architectures that reduce a model’s cultural surface area, grounding outputs in local context rather than relying on what the model internalized during pre-training. And it means being honest about what open-source can and cannot do.
Open-source gives you control over where the model runs. It does not give you control over what the model believes.
References
- Naous, T. & Xu, W. (2025). On The Origin of Cultural Biases in Language Models: From Pre-training Data to Linguistic Phenomena. Proceedings of NAACL 2025. https://aclanthology.org/2025.naacl-long.326/
- Rystrøm, J., Kirk, H. R., & Hale, S. (2025). Multilingual ≠ Multicultural: Evaluating Gaps Between Multilingual Capabilities and Cultural Alignment in LLMs. arXiv preprint arXiv:2502.16534. https://arxiv.org/abs/2502.16534