AI tools Arabic UAE test
Ask ChatGPT or Gemini to write you an email in English and it is flawless. Ask them to handle real Arabic, the kind people actually use in the UAE, and the cracks show fast.
We tested the big tools on Emirati dialect, Hijri dates, UAE addresses, and local context. Almost nobody benchmarks this seriously for the UAE, so here is the honest report.
First, the Good News. Formal Arabic Is Mostly Fine.
For Modern Standard Arabic, the formal written Arabic of news, contracts, and official letters, the big models are genuinely good now.
On the main academic test of Arabic knowledge, ArabicMMLU, GPT-4 scored around 72.5 percent, ahead of even strong Arabic-specific models. So if you need a polished formal press release or an official letter in standard Arabic, ChatGPT and Gemini will mostly serve you well.
That is the easy half. Now the half that matters in daily life here.
Where It Breaks 1: Khaleeji and Emirati Dialect
Real life in the UAE is not Modern Standard Arabic. People speak and text in Khaleeji, the Gulf dialect, with Emirati specifics. This is where the tools stumble.
Benchmarks built specifically for dialect, like AraDiCE and DialectalArabicMMLU, show a consistent pattern. The big multilingual models handle formal Arabic far better than they handle dialect, and dialect identification, generation, and translation all stay weak.
In practice, ask for a casual WhatsApp reply in Emirati dialect and you often get stiff standard Arabic back, or a vaguely Levantine or Egyptian flavour, because those dialects dominate the training data. It reads foreign to a local ear, the same way machine-translated English reads off to a native speaker.
Where It Breaks 2: Hijri Dates
Ask any of these tools to convert a Hijri date to Gregorian, or to tell you today’s Hijri date, and watch them drift.
Calendar conversion is a known weak spot for these models. They will often be off by a day or more, or confidently hand you the wrong year. The answer looks authoritative and is simply wrong.
This matters here more than almost anywhere. Official UAE documents, visa dates, Ramadan and Eid timing, and government deadlines all touch the Hijri calendar. Trusting an AI’s conversion without checking is exactly how you miss a date that matters.
Where It Breaks 3: UAE Addresses and the Makani Problem
The tools assume Western addressing. Street number, street name, postal code. The UAE largely does not work that way.
There are no conventional postal codes for everyday delivery, and locations are given by Makani number, by landmark, by building name, and by area. Ask an AI to format a UAE address properly and it will often invent a postal code or force the address into a US or UK template that no courier here would actually use. It does not understand Makani at all.
Where It Breaks 4: Local Context and Names
Per-country benchmark data shows UAE-specific knowledge is one of the weaker areas, even for Arabic-tuned models. Local holidays, institutions, and cultural specifics get fuzzy or simply wrong.
Name transliteration is inconsistent. Arabic letters like ain, haa, and khaa have no clean English equivalent, and the tools will spell the same Emirati name two or three different ways inside one document.
Mixed Arabic-English text, and Arabizi, the habit of typing Arabic in Latin letters with numbers, also confuse the tools on both formatting and meaning.
The Local Angle: The UAE Built Its Own
This gap is exactly why the UAE invested in its own models.
Jais, built by Inception, part of G42, with the university MBZUAI, is named after Jebel Jais, the country’s highest peak. Falcon, from Abu Dhabi’s Technology Innovation Institute, is another.
On dialect and regional tasks, Arabic-first models like Jais tend to beat the big multilingual ones, even though the global models still win on raw size and on English. That is the trade. Local nuance on one side, general power on the other.
How to Use the Big Tools for Arabic Without Getting Burned
Use them for formal Arabic, not dialect. For standard written Arabic, fine. For Emirati dialect, expect to fix it by hand, or reach for an Arabic-first tool.
Never trust an AI Hijri date. Always check it against an official source or a proper converter before you act on it.
Do not let them format UAE addresses. Do that yourself. They do not understand Makani.
Check every name transliteration for consistency across a document before it goes out.
For anything official or legal in Arabic, have a human who knows the dialect and the local context read it first. The tool is a draft, not a final.
The Bottom Line
The big AI tools have basically solved formal written Arabic. That is real progress, and worth using.
But the Arabic people actually live in here, the dialect, the dates, the addresses, the local detail, is still where they break. They were built for English first and the rest of the world second, and the UAE sits right in that gap.
Use them with your eyes open. Lean on them for the formal work, double-check the local work, and remember that the models built here, like Jais and Falcon, often understand here better than the famous ones do.
Sources
• ArabicMMLU: Arabic language understanding benchmark — https://arxiv.org/pdf/2402.12840
• AraDiCE: dialectal and cultural benchmarks for LLMs — https://arxiv.org/pdf/2409.11404
• DialectalArabicMMLU: dialect capabilities benchmark — https://arxiv.org/pdf/2510.27543
• Jais Arabic model (Inception / G42 / MBZUAI) — https://inceptionai.ai/jais/
• Falcon models (Technology Innovation Institute, Abu Dhabi) — https://falconllm.tii.ae/
Robius.news — Dubai, UAE — 2026 | Built to be first. Built to be trusted.
