54east / Insights / Perspectives

Perspectives · Whitepaper

Arabic isn’t a localization checkbox

Official Arabic is a register, structure, and archive problem. A language pack does not solve it.

54east · Abu Dhabi · Updated 2026-08-28 · Open access. Cited. No invented case studies.

Vendors selling AI document tools into UAE government treat Arabic the way they treat French or Japanese: a language pack. Translate the interface. Run a model trained mostly on English — and on whatever Arabic it saw on the open web — and ship it as “Arabic support.”

That is enough for a consumer chatbot. It is not enough for a ministerial letter.

The public record does not show ministries walking away from Arabic AI. It shows the opposite. In April 2025 the UAE Cabinet approved a Regulatory Intelligence ecosystem that uses AI for legislative research, drafting, evaluation, and enactment, with human insight still in the loop.¹ In July 2026 the Abu Dhabi Judicial Department announced it would implement an AI judicial platform under full human supervision. Gulf News launch coverage includes support for legal memoranda and judicial documents.² The live question is not whether to put AI on official Arabic. It is whether the system is built for the register ministries actually write.

This paper is about that gap: what official correspondence requires, what a generic “Arabic” model misses, and how to test a vendor on the institution’s own archive before procurement.

Three registers, not two

Arabic NLP has always been diglossic. Modern Standard Arabic (MSA) is the written variety of news, education, and pan-Arab media. Dialectal Arabic is everyday speech — and, heavily, social media.

Web-scale Arabic used to pretrain language models is not “overwhelmingly informal or dialectal.” It is heavily MSA news.

AraBERT was trained on Arabic Wikipedia, the OSIAN news corpus, and a 1.5-billion-word news collection.³ ARBERT was trained on 61 GB of MSA — newswire, Wikipedia, books, and the MSA slice of OSCAR. In that mix the OSCAR MSA portion is 31 GB; the Egyptian-Arabic slice is 32 MB.⁴ Common-crawl Arabic is a news-and-Wikipedia problem, not a dialect problem.

Dialect dominates social text. That is why MARBERT was pretrained on a billion Arabic tweets, and why NADI 2020 was built on Twitter.⁵ Even there the split is not clean: NADI’s organizers estimated roughly half their tweets as MSA.⁶ Newspaper comment threads sit in between — Zaidan and Callison-Burch’s Arabic Online Commentary labels found dialect in about two in five annotated sentences.⁷

Official correspondence is a third register. Neither news MSA nor tweets cover it well.

A ministerial letter is not a news article on letterhead. It has rank-tied honorifics, formulaic openings and closings, and a document grammar — the administrative letter, risāla idāriyya — that runs basmala, date, addressee, greeting, body, closing, signature, attachments.⁸ Industry Arabic records the practical cost: a tender cover letter that used the wrong form of address was treated as inappropriate and not submitted to the minister. Ḥaḍrat al-sayyid… al-muḥtaram is not “Dear Mr.” Closings such as wa-tafaḍḍalū bi-qabūl fā’iq al-iḥtirām are convention, not style.⁹ Sharjah’s Arabic Language Academy still trains government staff, in 2025, to distinguish official from informal correspondence. That skill is taught because the register is institutional.¹⁰

Grammatically correct Arabic that misses honorific, open/close, or risāla structure reads as fluent and untrained. For a citizen-facing notice, that is a style issue. For inter-ministerial or government-to-government correspondence, it reads as a lack of institutional seriousness. The document comes back for rewrite. The AI has added a step.

Structure is not a translation leftover

A system that translates an English memo workflow into Arabic keeps English structural logic — subject, body, conclusion — and swaps the words. The result is Arabic script on an English memo.

Official Arabic correspondence has its own document shape. Requests, notifications, and approvals are not the same letter with a different verb. How a document names the prior correspondence it answers, who is asked to act, and on whose authority, is part of the document’s function.

Do not confuse that with the routing slip. In most UAE institutions the approval chain lives in the correspondence-management system — tawqīʿ, workflow, the CMS — not only in the prose. The letter still has to carry a reconstructable record of what is being requested, from whom, and on whose prior authority. If the generated text obscures that, routing fails even when the workflow engine is correct.

Volume is a review problem

A single well-drafted Arabic document from a general-purpose model, edited by a trained officer before it goes out, is usually fine. The officer catches register and structure.

A correspondence system exists to cut that review burden at institutional volume. If every output still needs a fluent, trained reader to catch register and structure, drafting time has moved into review time. That is not a workload reduction.

The gap is between “the model can produce Arabic” and “the system can be trusted to produce official Arabic with a proportionate review burden.” The first is table stakes. The second is built around the institution’s conventions — register, document types, approval chains, archive — not around the general capability of the underlying model.

What the system has to do

Four capabilities separate a system that reduces work from one that produces plausible Arabic for full re-review.

Intake and classification on the entity’s actual document types. A ministry’s incoming mail has entity-specific categories. A generic classifier trained on public documents will not see them.

Routing that matches the real approval chain. Every institution’s sign-off structure is its own. A template chain creates rerouting.

Drafting in the institution’s own register, learned from its own archive. The reliable way to get honorifics, open/close, and request-versus-notification framing right is to build drafting patterns from the institution’s historical correspondence — not a generic formal-Arabic set.

Archive in the language it was signed in. A document generated or translated for internal processing, then archived in a different form, is a record that does not match what was signed and sent. The system of record stays the system of record. The trail has to be reconstructable in the exact language and form that existed at approval.

The cost is reputational, then operational

Documents that go external — another ministry, a citizen, a counterpart government — and read as competent but off in register or structure reflect on the sending institution. That is a different risk than an awkward chatbot reply. Procurement underweights it, because the demo is one well-chosen example. Register and structure fail on real institutional documents, at real volume.

Test it on your own archive

Do not buy on the vendor demo.

Give the vendor a sample of the institution’s own previously sent correspondence — sensitive specifics redacted — spanning a few document types and approval chains. Ask the system to draft equivalent new documents. Score the outputs the way the institution’s own correspondence staff would: register, structure, routing accuracy. Not generic Arabic fluency.

That test surfaces the gap between “handles Arabic” and “handles our correspondence.” It also puts the people who will live with the system into the procurement decision. Adoption is better when the operators validated the documents.

Bilingual record, one system of record

Gulf institutions often need an Arabic original and an English working translation, especially where staff or counterparts work primarily in English. Generating one and bolting a machine translation onto the other creates a dual-record problem: which version was signed, and does the secondary version drift after successive edits?

Treat this as architecture, not an afterthought. One designated version — typically the Arabic original for official government correspondence — is the system of record. Any translation is marked as a working translation, not an equivalent authoritative copy. If the original is amended after first draft, the translation is resynchronized, or it is marked stale. Leaving this implicit is how an institution later cannot say which version reflects what was approved.

This is operational practice, not a statute. It is still how a bilingual archive stays reconstructable.

Close

Arabic correspondence at government volume is not a translation problem with an AI layer on top. It is a system-design problem — register, structure, approval-chain logic, archival integrity — that a model’s general Arabic fluency does not solve.

Ask to see the system work on your document types and your approval chains before treating “handles Arabic” as “handles our correspondence.”

54east builds correspondence systems designed in Arabic from the institution’s own archive and approval structure — not translated from an English template. Register and approval chains, not a language pack. In-country. Owned by the institution that signed. If review is still the bottleneck, start with a briefing.

---

Notes

1. UAE Cabinet, 14 Apr 2025: Regulatory Intelligence ecosystem and Regulatory Intelligence Office; AI tools for research, drafting, evaluation, and enactment. WAM.
2. Abu Dhabi Judicial Department, 28 Jul 2026: WAM reports coordination meetings to implement an AI judicial platform under full human supervision; first phase expected September 2026, 18-month rollout. WAM. Launch coverage listing support for drafting legal memoranda and judicial documents (not WAM): Gulf News.
3. Wissam Antoun, Fady Baly, and Hazem Hajj, “AraBERT: Transformer-based Model for Arabic Language Understanding,” OSACT 2020. Pretraining data: Arabic Wikipedia, OSIAN news, 1.5-billion-word news corpus. ACL Anthology.
4. Muhammad Abdul-Mageed, AbdelRahim Elmadany, and El Moatez Billah Nagoudi, “ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic,” ACL 2021. ARBERT: 61 GB MSA (OSCAR-MSA 31 GB; OSCAR-Egyptian 32 MB). MARBERT: ~1 billion Arabic tweets. ACL Anthology; arXiv; GitHub. OSCAR (Open Super-large Crawled Aggregated coRpus): oscar-corpus.com.
5. Ibid. MARBERT trained on a random sample of 1 billion Arabic tweets (128 GB / 15.6 billion tokens).
6. Muhammad Abdul-Mageed, Chiyu Zhang, Houda Bouamor, and Nizar Habash, “NADI 2020: The First Nuanced Arabic Dialect Identification Shared Task,” WANLP 2020. In-house MSA–DA model estimated 49.5% / 46.6% / 49.7% MSA on train / dev / test tweets. ACL Anthology; arXiv.
7. Omar F. Zaidan and Chris Callison-Burch, “The Arabic Online Commentary Dataset,” ACL 2011. 108k labeled sentences; 41% labeled as having dialectal content. ACL Anthology.
8. Sayidolim Rayimjonov, “Methodology of Official Correspondence in Arabic: Theory and Practice,” Journal of Interdisciplinary Human Studies 2, no. 3 (2026). Risāla idāriyya structure: basmala, date, addressee, greeting, body, closing, signature, attachments. DOI.
9. Industry Arabic, “Let’s Address the Address: A Guide to Formal Letters in Arabic,” 18 Feb 2023. Wrong form of address; rank titles; formulaic openings and closings. industryarabic.com.
10. Arabic Language Academy in Sharjah with Sharjah Directorate of Human Resources, course “Official Correspondence: Formal and Informal Communication,” 14 May 2025. Sharjah24.

Related whitepapers