Who Teaches the Machine Our Language and Values? AI and Cultural Identity

A model that speaks in your tongue and is at home with what others are at home with

Research on cultural bias in language models has observed a recurring pattern: a model asked in an Arab-Islamic context that calls up something familiar from another culture. A model learns not language alone but, with it, what its writers find familiar.

Figures on presence

Common Crawl publishes statistics on the primary language of the web pages it collects. In the latest crawl at the time of writing (CC-MAIN-2026-43), English made up about 42.6% of pages and Arabic about 0.67%. This is not the training proportion in any given model, since companies select and balance their data, but it is a reminder that far less is available in Arabic than in English.

What have the studies measured?

• Naous and colleagues: they evaluated 16 models on Arabic tasks (story generation, named-entity recognition and sentiment analysis), using the CAMeL collection of 628 prompts and 20,368 entities of eight types that contrast Arab and Western culture. They found a lean towards entities associated with Western culture in multilingual and Arabic models alike, and cases of stereotyping and bias. They concluded that common sources such as Wikipedia may not suffice on their own to build culturally aware models.
• Tao and colleagues: they compared five OpenAI models with national value surveys and found their values closest to English-speaking countries and Protestant Europe. When the model was asked to answer from the perspective of a particular country ("cultural prompting"), alignment improved in 71% to 81% of countries and territories for the newer GPT-4 family models.

These are results for particular models at a particular time, so they should not be generalised to every model or every question, but they are enough to alert the user. They also carry a practical benefit: telling the model the context you want, your country, your customs and what you observe, brings it closer to what you intend, though it does not guarantee it will get it right every time. And what touches religion and morals in particular should not be taken from a model without review.

What UNESCO said

In November 2021 UNESCO adopted a Recommendation on the Ethics of Artificial Intelligence, endorsed by 193 countries, which made among its first values "ensuring diversity and inclusiveness" and counted the creation of new means to safeguard cultural diversity among AI's hoped-for benefits. As for Kate Crawford, her book Atlas of AI treats the power, politics and planetary costs of artificial intelligence, an angle that reminds us these systems are products with owners and interests, not a neutral mirror.

Not isolation

This does not mean rejecting the technology, for civilisations take from what others have while keeping what is theirs. What is required is that we sit at the table where it is made, not only in the seat where it is consumed:

• Authentic data: digitising good Arabic texts and organising them under the supervision of specialists in the language.
• Local evaluation: measures that test cultural understanding, not merely soundness of phrasing.
• Specialist review: matters touching religion and values are reviewed by scholars before relying on a machine's output.
• Informed prompting: writing the request with its cultural and local context improves the result, but does not replace checking.
• User awareness: a machine's output is an opinion open to question, not a judgement immune from error.

For a comparison of Arabic models and the sovereignty of production, see the article "Arabic Artificial Intelligence".

A person is entrusted with his tongue and his values as he is entrusted with his wealth; whoever teaches the machine his language faithfully, and reviews what it says with a critical eye, preserves what is his and benefits from what others have.