Check out this story in the debut Godmother print zine Issue 0.

The race to develop sovereign LLMs is growing with a sense of urgency and is emerging as a crucial topic for the coming century. 

With English dominating most of the widely used LLMs (ChatGPT, Claude, Mistral, Llama, and Gemini), context and language are both crucial. When prompts generate responses that are based on Western concepts, the risk of identity loss increases due to false or misleading information, and this is where context engineering matters. Contextually rich frameworks lay the ground for deeper understanding and evaluation, contributing towards addressing challenges like bias, security, and transparency. Since LLMs generate different answers to the same question, each model has its own version of the story to present, making LLM security a concern. At a time when geopolitical tensions threaten world peace, sovereign AI infrastructure is a matter of urgency that can’t be ignored. 

What is LLM Sovereignty? 

“LLM sovereignty is really about having control over the AI models we use: how they are trained, what data they learn from, which languages they support, and who defines the rules around them. So it is not just a technical issue. It is also about trust, representation, and making sure these systems reflect real needs and priorities,” says Dr. Feda Almuhisen, an AI research engineer at CEA. 

Trust and representation are foundational to identity – whether for a nation or an individual. In the age of algorithms, sovereign LLMs are a powerful tool because they preserve linguistic diversity and shape how cultures preserve and define themselves in the future. While there are over 166 LLM models available, the biases encountered with them are notable. 

Last year, when China released DeepSeek-R1 offering high performance at lower costs, there were questions about its efficiency, even though it rivalled models from OpenAI, Anthropic, and Google. At the time, a study by US-based Enkrypt AI revealed that 83% of bias tests on DeepSeek-R1 detected discriminatory output, with biases noted across race, gender, health and religion. The study also found that 78% of cybersecurity tests had tricked R1 into generating insecure or malicious codes, leading experts to raise a red flag on how technology could reflect a country’s geopolitical strategy, and, conclusively, impact national security concerns if caution is not extended. 

DeepSeek’s latest model, V4 – released this year, is also open source and available to anyone, unlike the American models, showing how far and wide a reach it has over the global population. The risk, then, is not just fundamental autonomy, but governance at the highest level across all sectors, with language, race, and stereotypes adding a layer that must be tackled with judiciousness when it comes to the Arab region. “From an Arab perspective, this is important, because most language models today are still built mainly around English and on data that comes from outside the region,” points Almuhisen

Language, Culture, and National Identity

Not merely a tool for expression, LLMs and linguistic identity are at the heart of the global conversation on digital sovereignty because of its impact on behaviour, and the reaction it elicits. Like in a game of Chinese whispers, what is said in one language can be misinterpreted in another, causing a chaotic result if it is not communicated accurately. 

Similarly, if an LLM is trained on data sets that do not include accurate historical data, cultural nuances, and dialects, it is more likely to respond with unreliable, biased, offensive, or incorrect responses. This is where stereotyping propaganda can negatively influence people’s behaviour. 

A study from Johns Hopkins found that multilingual LLMs showed preference towards high-resource languages such as English, Chinese, and German, as compared to low resource languages like Hindi or Arabic. Their study revealed how American perspectives are often forced onto speakers of low-resource languages, including biased information on the Israel-Gaza and Russian-Ukraine wars. As part of the experiment they conducted, LLMs were fed information based on a different set of data, one with truthful information available, and the other with alternative and conflicting information. The resulting outputs from the queries showed biased responses on the international conflicts, revealing how people can walk away with different understandings. 

But even if a language is widely spoken, various dialects may still be misrepresented or excluded, shares Almuhisen. “As a native Arabic speaker, I noticed and encountered these limits when using models available today like ChatGPT. They still need to be better adapted to Arabic linguistic and cultural contexts,” she says. “Even when Arabic is included, the data available online does not tell the full picture. It often misses a lot of the richness of the language, the diversity of dialects, and the realities of different Arab communities and cultures.” 

“Dialects matter because if we ignore them, we end up building models that may look strong on paper but do not serve people well in practice,” Almuhisen adds. “Recent work also shows that models often understand dialectal Arabic better than they generate it, and that some training choices can push them back toward Modern Standard Arabic instead of reflecting dialect use more naturally. That means training a strong Arabic LLM is not just about adding more Arabic text, but about making sure the data reflects how Arabic is actually used across different communities.”

Inter-Connectedness & Dialect Diversity

Dialect diversity in the Arabic language is of relevance to the Middle East because, even though there is Modern Standardized Arabic, people speak dialects which vary in vocabulary, expression, and structure. While LLMs are trained on data sets based on articles, books, and Wikipedia entries, they are designed to produce outputs based on what is already fed into the system. The simple understanding is: what goes in, comes out. The lesser the training data, the more limited the output, but including broader parameters is not the only concern. “Dialects matter because if we ignore them, we end up building models that may look strong on paper but do not serve people well in practice,” shares Almuhisen

In terms of the Arabic language, while it may perform well on benchmarks, it may not reflect how people actually speak and write, she adds. “Recent work also shows that models often understand dialectal Arabic better than they generate it, and that some training choices can push them back toward Modern Standard Arabic instead of reflecting dialect use more naturally.” Another point she makes is that Arabic online content is not only diverse in dialect, but often mixed with other languages or written in Arabizi, which adds another layer of complexity for language models. “That means training a strong Arabic LLM is not just about adding more Arabic text, but about making sure the data reflects how Arabic is actually used across different communities.” 

The Women Shaping the Dialogue 

To counter this, Arab women are slowly stepping up to shape the dialogue on LLM sovereignty. A few of them are paving the way for younger women, because of the many implications that it will have on future generations.

As a woman of Arabic origin leading the scientific field, Almuhisen recalls how she has always found herself in a male-dominated environment, and appreciates women who are raising the bar on LLM sovereignty: “At the 2025 Doha conference on AI and the characteristics of the Arabic language, notable contributions were made from women as such as Hend Al-Khalifa of King Saud University on Arabic language modelling, and Rafeef Al-Sayyid on Arabic AI policy.”

“In language technologies, Karen Spärck Jones is a major name because of her important work in natural language processing and information retrieval. And in today’s AI landscape, Fei-Fei Li is one of the clearest examples of someone who has shaped both the science and the human centered vision of the field,” she shares.

These women are inspirational for their contributions to research, and with countries now prioritizing digital sovereignty, there is a new shift in powers. It is no longer a matter of technological advancements, it’s about the preservation and protection of democratic values, identity, and culture.    

Karoline M. is an editor, writer and strategist based in Dubai.