Arabic-First AI: Making AI Content Work for Gulf Audiences

Ask a general-purpose AI tool for a product description, a social post or a newsletter in Arabic and the output usually has the same tell: it is grammatically fine but reads like it was thought in English and converted. The rhythm is off, the register is neither properly formal nor properly local, and a Gulf reader clocks it in a sentence. For a business in Riyadh, Dubai or Doha that wants its Arabic content to carry the same weight as its English, that is a real problem, not a cosmetic one.

The gap is not random. It comes from how these models are built, and it is fixable with the right process around the tool.

Why generic models are weak in Arabic

Large models learn from the web, and the web is not balanced. The team behind Qatar’s Fanar model (2025) cites Common Crawl figures putting English at about 46% of all text on the web, while Arabic, spoken by more than 400 million people, sits at around 0.5% (Fanar). A model trained on that mix has seen far less Arabic, and far less variety of Arabic, than it has English.

The scale of the shortage shows up in what it takes to build an Arabic model. Saudi Arabia’s ALLaM (2024), developed by the National Center for AI at SDAIA, was trained on 540 billion Arabic tokens, but half of that, 270 billion, was English text machine-translated into Arabic because there was not enough natural Arabic to reach the target (ALLaM). When even a national Arabic-first project has to pad its corpus with translations, a general model with no Arabic focus is thinner still.

This is why the output feels translated. In effect, some of it is: the model learned part of its Arabic from translated English in the first place.

The dialect question you have to answer

Arabic is not one register. Modern Standard Arabic (MSA) is the formal written language, shared across the Arab world, used in news, government and formal marketing. Gulf dialect, Khaleeji, is what people actually speak in Saudi Arabia, the UAE, Qatar, Kuwait, Bahrain and Oman, and increasingly what they expect in casual brand communication, chat support and short social copy.

Models handle this badly by default. The ALLaM team notes that identifying dialectal Arabic in training data is still an open research problem, so they did not deliberately include dialect datasets at all (ALLaM). Fanar is one of the few that explicitly supports Gulf, Levantine and Egyptian dialects. A generic model asked for casual Arabic will usually give you MSA with a few colloquial words dropped in, which reads neither here nor there.

So the choice is yours to make, per use:

• MSA for formal content. legal pages, corporate announcements, long-form thought leadership, anything pan-Arab.
• Gulf dialect for conversational content. social captions, SMS and WhatsApp, chat support, short ads aimed at a local audience.
• One register per piece. a paragraph that drifts between registers is the clearest sign of machine output.

The hidden tax: tokens, cost and truncation

Arabic is also more expensive to process. A study presented at NeurIPS 2023 found that the tokenizer used by ChatGPT and GPT-4 needs about three times as many tokens to encode Arabic as the same text in English (tokenizer study). The knock-on effects are practical: you pay per token, so Arabic generation and Arabic context both cost more; latency rises with token count; and a long Arabic brief or document fills a fixed context window roughly three times faster than its English equivalent, so more of it gets truncated or forgotten.

This is part of why teams quietly draft in English and translate: it is cheaper and the tools behave better. It is also why that shortcut produces weak Arabic. If you are going to work in Arabic, budget for it costing more per word and keep prompts and reference documents tighter than you would in English.

What Arabic-first actually means

Arabic-first is not a translation step bolted onto an English workflow. It is a small set of habits that change where Arabic sits in the process.

1. Brief in Arabic, or bilingually. if the instruction to the model is in English, the output inherits English structure. A short Arabic brief, even a rough one, pulls the draft toward Arabic phrasing.
2. Keep a bilingual brand glossary. fix the Arabic for your product names, key features, tagline and tone words, and feed it to the model every time. Without it, the same term comes out three different ways across a week of content.
3. Edit, do not proofread. a native Gulf speaker should rework the draft for rhythm and register, not just check spelling. Budget an hour of real editing per substantial piece, not five minutes.
4. Decide the register before you generate. MSA or dialect, chosen for the audience and the channel, and stated in the brief.
5. Localize the specifics. currency, dates, names, examples and cultural references, not just the words.

This mirrors what Nielsen Norman Group has found across years of testing international sites: content should be genuinely localized, not translated verbatim, because product descriptions, thought leadership and domain jargon need adapting to local language patterns, and tone of voice varies by region in ways a literal translation misses (Nielsen Norman Group).

Checklist for the Arabic-first workflow:

1. Register chosen. MSA or Gulf dialect, decided before you generate.
2. Brief written in Arabic. even a rough one, so the draft inherits Arabic structure.
3. Glossary loaded. the bilingual brand terms fed to the model every time.
4. Specifics localized. currency, dates, names and cultural references, not just words.
5. Native editor pass. a Gulf speaker reworks rhythm and register, not spelling.
6. Right-to-left QA. the layout mirrors and the Arabic has room to run longer.

Tooling and layout for a Gulf SME

You have three broad options, and most small teams end up combining them.

• A frontier model with strong scaffolding. the general tools are usable for Arabic if you brief in Arabic, supply a glossary and edit properly. For many SMEs this is enough, and it is the least operational overhead.
• A region-built model. Fanar (Qatar), Jais (UAE) and ALLaM (Saudi Arabia) are trained with Arabic at the centre and handle register and cultural context better out of the box. Worth testing if Arabic is more than half your output.
• A hybrid. draft with whichever model reads best in Arabic for that task, then run a consistent human edit and glossary pass regardless of source.

Layout is the other half. Arabic runs right to left, and Nielsen Norman Group’s usability testing found Arabic text also tends to run longer than English, so interfaces should allow at least 50% more room for text and support a mirrored reading pattern where the eye scans down the right side of the page (Nielsen Norman Group), with the whole layout mirrored, navigation and alignment included, not just the text. Language options should be labelled in words, Arabic and English, not flags, since one flag cannot represent a language spoken across more than twenty countries.

Why this is worth the effort

The commercial case is straightforward in this region. PwC estimates AI will add about US$320 billion to the Middle East economy by 2030, with Saudi Arabia at roughly US$135 billion and the UAE at US$96 billion (PwC). The businesses that capture a share of that will be the ones whose Arabic customer experience is as strong as their English, not an afterthought bolted on by a translation plugin.

Arabic-first content is not about using a different tool. It is about treating Arabic as a language you write in, with a brief, a glossary, a chosen register and a real editor, rather than a language you convert into at the end.

Mistakes to avoid

• Writing in English and translating at the end. the fastest route to content that reads foreign to a Gulf audience.
• Letting the model pick the register. you will get a flat MSA-and-dialect hybrid that suits no channel.
• Proofreading instead of editing. spelling is not the problem, rhythm and register are.
• Reusing English structure and length. Arabic sentences and layouts need room to breathe, on the page and in the design.
• A one-off glossary. if it is not reused on every piece, terminology drifts within days.

If you want help building an Arabic-first content process, or choosing between a frontier model and a region-built one for your mix, that is the kind of work we do at Gaveau Strategy.