AI Governance in the Arab Region
Last updated:
AI governance in the Arab region differs from the European or North American debate in three ways: Arabic is under-served by models trained overwhelmingly on English, capability is distributed very unevenly between Gulf states and conflict-affected economies, and much regional AI adoption happens through procurement rather than domestic development. Governing procured AI in a second language is the region\’s defining problem.
Artificial intelligence governance in the Arab region, written from Beirut. The regional debate is not a translation of the European one: Arabic is structurally disadvantaged in models trained mostly on English, national capability varies enormously, and most adoption arrives through procurement rather than domestic development.
Why the regional context is genuinely different
Three conditions separate AI governance in the Arab region from the debate that dominates English-language policy writing.
Language. Models are trained overwhelmingly on English. Arabic is not a minor language, but it is under-represented relative to its number of speakers, and it is not one language in practice: Modern Standard Arabic is what models mostly see, while people actually write and speak in Levantine, Gulf, Egyptian, and Maghrebi varieties. A system evaluated only on Modern Standard Arabic will report accuracy it does not have in the field. This is the argument set out in When 95% of AI’s Brain is English, the Rest of the World Pays a Tax.
Uneven capability. The region spans states with substantial sovereign compute ambitions and states whose public administrations are rebuilding after conflict. A single regional AI policy prescription does not fit both. What travels between them is not a strategy document but a method for deciding what must be under national control and what can be partnered.
Procurement, not construction. Most AI entering Arab public administrations is bought rather than built. That makes vendor assurance, data residency, and the right to inspect and withdraw a system the central governance questions, ahead of model development ethics.
Governing AI you bought rather than built
When a system is procured, the governing instrument is the contract and the assurance process around it, not an internal model card. Four questions decide most of it. Where does the data go, under whose law, and can inference be kept inside the country? What can be inspected: does the buyer get evaluation results on their own data and languages, or only the vendor’s benchmark? What happens on failure: who is accountable, and is there a defined condition under which the system is withdrawn? And what happens at exit: can the institution leave with its data and its records intact?
The vendor assurance question set in the AI security and assurance playbook is written for exactly this situation, and the sovereignty layer is treated in full on the sovereign AI strategy page.
Evaluating Arabic AI honestly
The practical rule is to test dialect by dialect against the population the system will actually serve, using text and speech collected from that population rather than a translated benchmark. Accuracy figures reported on Modern Standard Arabic should be treated as an upper bound, not an expectation. Right-to-left rendering, mixed-script input, and transliterated Arabic written in Latin characters all need to appear in the test set, because they appear in real messages.
Where a system serves displaced or vulnerable people, language coverage is a protection question rather than a quality metric: a person who cannot be understood by the system cannot access the service behind it. That is the connection between this cluster and the Last-Mile AI Framework, which treats low-resource languages as a first-class design input.
Where this perspective comes from
Shahzad Asghar is Head of Data and Digital Solutions at UNESCWA, the United Nations Economic and Social Commission for Western Asia, the UN regional commission serving the Arab States, based in Beirut. The role covers data engineering, web delivery, infrastructure, and cybersecurity, including a digital discovery that inventoried 181 databases and 145 portals across the Commission with lifecycle verdicts. Earlier work included designing and delivering institutional cybersecurity governance and incident response training for IT and security professionals across the MENA region, and building refugee-facing systems in Jordan.
Related reading: AI governance in the United Nations, AI leadership and governance, and the playbooks. The United Nations body now producing the shared evidence base for these debates is explained on the UN Independent International Scientific Panel on AI.
Common questions
What is the state of AI governance in the Arab region?
Three conditions define it. Arabic is under-represented in models trained overwhelmingly on English, so systems perform worse than their reported benchmarks suggest. National capability varies enormously, from states with sovereign compute ambitions to administrations rebuilding after conflict, so a single regional prescription does not fit. And most AI entering Arab public administrations is procured rather than built, which makes vendor assurance, data residency, and the right to inspect and withdraw a system the central governance questions.
How do you build AI for Arabic language and dialects?
Treat Arabic as several languages rather than one. Models mostly see Modern Standard Arabic, while people write and speak Levantine, Gulf, Egyptian, and Maghrebi varieties. Test dialect by dialect against the population the system will actually serve, using text and speech collected from that population rather than a translated benchmark. Include right-to-left rendering, mixed-script input, and transliterated Arabic written in Latin characters, because all three appear in real messages.
Why does Arabic AI underperform English AI?
Because training data is overwhelmingly English, and because Arabic is diglossic in practice: the written standard that models see is not the variety most people use. A model evaluated on Modern Standard Arabic will report accuracy it does not achieve on Levantine or Maghrebi speech. Accuracy on Modern Standard Arabic should be read as an upper bound rather than an expectation.
How should Arab governments govern AI they buy rather than build?
Four questions decide most of it. Where does the data go, under whose law, and can inference be kept inside the country? What can be inspected: does the buyer receive evaluation results on their own data and languages, or only the vendor's benchmark? What happens on failure: who is accountable, and is there a defined condition under which the system is withdrawn? And what happens at exit: can the institution leave with its data and records intact? When a system is procured, the governing instrument is the contract and the assurance process around it.
Who works on Arabic-language AI governance?
Shahzad Asghar works on AI governance in the Arab region as Head of Data and Digital Solutions at UNESCWA, the United Nations Economic and Social Commission for Western Asia, the UN regional commission serving the Arab States, based in Beirut. His work covers data engineering, infrastructure and cybersecurity for the Commission, institutional cybersecurity training delivered across the MENA region, and refugee-facing AI systems built in Jordan.
Get the next essay by email
One practical essay a month on AI governance, agentic AI, and digital delivery in the UN system. No marketing, no forwarding of your address.