Localising AI: Cultural Context, Not Interface Translation

Ask a model to complete an Arabic sentence about food, clothing or a social custom, and it may choose entities belonging to another culture while writing impeccable Arabic. This is what researchers tried to measure precisely. So is it enough to Arabise the interface to say we have localised artificial intelligence?

What the research says

Tarek Naous and colleagues tested sixteen models on generating Arabic stories, recognising entities and sentiment analysis, using CAMeL, a set they built: 628 natural prompts and more than 20,000 entities of eight types, contrasting Arab and Western cultures. They concluded that models, multilingual and Arabic-only alike, lean towards entities associated with Western culture, and show stereotyping and cultural unfairness, and that some common sources for training Arabic models, such as Wikipedia, may not on their own be suitable for building culturally aware models unless treated first. For the general treatment of this topic, see the article "Algorithmic Bias".

In another study (Cao and colleagues, 2023) ChatGPT was found to align strongly with American culture when asked in its context and to adapt less to others, and prompting in English reduced the diversity of answers and pulled them towards American norms. Language alone does not decide culture, but data and alignment do.

Why is translation not enough?

Because culture is not vocabulary. When a model is asked about a social, family or religious matter, it may present a view that grew up in another context as the natural answer. This is no call for isolation: global knowledge is a shared inheritance from which we take and benefit, but we ask that the tool understand us before it addresses us. The clearest example is matters of religion: the model is not a mufti, a fatwa belongs to its qualified scholars, and whatever is produced in this field is reviewed before anything is built on it.

What is happening in practice?

Efforts have begun to treat the matter at its roots. The authors of the ALLaM model (2024) describe expanding the model's vocabulary, training it on a mix of Arabic and English, and aligning it with human preferences, and they report leading results on Arabic tests such as Arabic MMLU, ACVA and Arabic Exams; these are results presented by its developers, so they should be read with caution until repeated in independent evaluation. SDAIA (the Saudi Data and Artificial Intelligence Authority) announced the model in May 2024 in cooperation with IBM on its watsonx platform. And Gulf News reported that the HUMAIN Chat app, launched in Saudi Arabia in 2025, runs on ALLaM 34B, and that HUMAIN is owned by the Public Investment Fund.

How do we measure localisation?

Measurement is what separates publicity from truth. In 2024 Koto and colleagues presented ArabicMMLU: about 14,575 questions in 40 tasks from school exams in North Africa, the Levant and the Gulf. They tested 35 models, and the best Arabic-oriented model did not exceed 62.3%; these are 2024 figures that may have changed since. The test's merit is that its questions come from Arab school exams, not from translations of foreign tests.

Five tasks that never end

1. Selected Arabic data: heritage and contemporary, cleaned, with documented sources and rights respected.
2. Local benchmarks that test cultural understanding, not translation alone.
3. Review by specialists on matters of religion and values, before launch and after.
4. Regular updating, since language and questions renew themselves.
5. Transparency about what the model was trained on and what it cannot do, so the user knows when to verify for themselves.

For knowledge sovereignty and the wider view, see the article "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models". Here the criterion is simpler: does the tool understand what we mean, not merely what we say? If we do our part in taking the means, may God grant us success in what befits our language and values.