Who Owns the Cultural Layer of Arab-Language AI?
Building better Arabic AI is not only a technical challenge. It raises commercial and cultural questions about data, dialects, creative rights and who gets to define reliable Arabic intelligence.
When an artificial-intelligence system responds in Arabic, its fluency can create an impression of cultural understanding.
The two are not the same.
A model may produce grammatically convincing Modern Standard Arabic while misunderstanding a regional expression, flattening the differences between Arab societies or reproducing assumptions learned from translated material. It may recognise a famous writer yet know little about local publishing, oral history or contemporary cultural life outside the region’s largest digital markets.
As Arab states and companies invest in language models, the central question is moving beyond whether AI can speak Arabic. It is becoming: whose Arabic, trained on whose material, governed by whom and for whose benefit?
Egypt brought that issue into sharper focus in February with the launch of Karnak, described by the government as a national large language model. Announced applications included an Arabic educational tutor covering language and Egyptian history and a legal and regulatory assistant for citizens and small businesses.
These are practical uses. They also show why the cultural layer of AI matters. A tutoring system influences how history is presented. A legal assistant must distinguish authoritative text from informal commentary. A public-facing model needs to understand the language people actually use without treating one dialect or social group as the default for the region.
Arabic is not a single dataset
Arabic combines a shared written language with wide dialect variation and distinct national, religious and cultural contexts. Much of the material available online is unevenly distributed across those contexts.
A 2025 research project involving 44 researchers built a culturally and linguistically diverse dataset covering all 22 Arab countries. Its evaluation found that leading models still had limitations and that Egypt and the UAE were better represented than countries including Iraq, Mauritania and Yemen.
That imbalance is not merely a technical inconvenience. It shapes which users receive accurate answers, whose expressions are recognised and which communities appear in the system’s implicit picture of Arab life.
The commercial opportunity is substantial. Better Arabic systems could improve education, customer service, search, accessibility, media production and access to government services. But a model trained mainly on easily available material will tend to reproduce the geography of the existing internet rather than the diversity of the language.
Correcting that requires deliberate investment in dialectal, historical and specialist data—and people capable of evaluating it.
Cultural material is also intellectual property
The training data required to improve Arabic AI may include books, journalism, films, music, photographs, archives and recorded speech. Much of that work was created by people and institutions that hold rights in it.
This creates a difficult economic question. If cultural material helps make a commercial model more capable, how should its creators and custodians participate in the value generated?
The answer cannot simply be that anything accessible online is available without meaningful terms. That approach risks transferring value away from already fragile publishing, media and cultural institutions into a small number of technology platforms.
It could also discourage archives and publishers from digitising material if they believe doing so makes it easier for others to extract value without permission or attribution.
Licensing will not be simple. Large archives contain layered and sometimes uncertain rights. Collective agreements, standard licences and trusted data intermediaries may become more practical than negotiating work by work. Public-domain material should be made genuinely accessible, while protected work requires transparent terms appropriate to its use.
The region has an opportunity to design these markets early rather than importing rules after commercial practices are entrenched.
National models need public accountability
A nationally backed language model can support local control over infrastructure and sensitive data. It can also provide a platform that domestic startups and public agencies build upon.
The label “national,” however, should bring obligations as well as prestige.
Users should know who developed the system, what broad categories of data shaped it, which institutions evaluate it and how errors in education, law or public services can be challenged. Independent researchers should have enough access to test accuracy, bias, dialect performance and security.
Claims of cultural alignment should be measurable. Does a model understand the same question across dialects? Does it confuse social norms between countries? Does it cite reliable sources when discussing law or history? Does performance deteriorate for places with fewer digital resources?
Research benchmarks are beginning to expose these gaps. Public procurement can accelerate that work by requiring clear evaluation before Arabic AI is deployed in sensitive settings.
Culture cannot be reduced to localisation
Technology companies often describe Arabic as a localisation task: translate the interface, adapt some examples and release a regional version.
Culture is deeper than interface language.
It appears in humour, metaphor, historical memory, family relationships, religious and secular traditions, visual symbols and the difference between formal and everyday speech. It also changes. No committee or dataset can define a single fixed set of Arab values without excluding meaningful disagreement.
The objective should therefore not be to build a model that claims to represent Arab culture as one thing. It should be to create systems capable of recognising difference, signalling uncertainty and allowing communities to contest errors.
UNESCO has warned that AI can concentrate cultural content, data, markets and income in a small number of actors, with consequences for linguistic diversity and cultural expression. Its global roadmap for multilingualism calls for safer, higher-quality tools for languages with limited digital resources.
For the Arab world, that is both a policy challenge and an industrial opening.
The institutions around the model will matter most
The long-term winners in Arabic AI may not be defined only by model size. They may be the organisations that assemble trusted data, secure legitimate rights, evaluate dialects properly and build applications people can rely upon.
That ecosystem should include universities, publishers, broadcasters, museums, libraries, archives, startups and independent researchers—not only government agencies and global technology firms.
Several practical steps would move it forward:
- transparent licensing frameworks for Arabic text, audio and visual collections;
- funding to digitise underrepresented archives with clear rights and public-interest access;
- benchmarks spanning countries, dialects and specialist domains;
- independent evaluation for systems used in education, law, health and government;
- ways for creators and institutions to opt in, negotiate terms or challenge misuse; and
- support for open research resources that smaller companies can build upon.
Arabic AI will not be culturally credible simply because it is developed in the region. Credibility will depend on how its knowledge is assembled, how openly it is tested and whether the people whose language and work give it value have a meaningful place in its economy.
The cultural layer is becoming infrastructure. The question of who owns and governs it should be settled before it disappears inside the model.
Sources
