Does the Machine Speak Arabic, or Merely Imitate It?

A richness that confuses the computer

Arabic is derived from roots that branch into dozens of forms, and it is usually written without short vowels. Strip a word of its diacritics and its possible morphological readings remain many. A research paper on diacritisation and morphological analysis cites an estimate, taken from Nizar Habash (2010), that an Arabic word has about 12 analyses on average. This is an estimate, not a direct measurement made by the paper itself.

Add to this that "Arabic" is a spectrum running from the written standard language to spoken dialects. A model that has mastered the language of newspapers may not handle the language of the home and the market, because dialects are spoken more than written, so less of them is available for training.

The data gap

Common Crawl, the organisation that regularly gathers the web, publishes statistics on the primary language of its pages. In the latest crawl at the time of writing (CC-MAIN-2026-43), English made up about 42.6% of pages and Arabic about 0.67%. These are not the training proportions of any given model, but they point to the imbalance in the two languages' presence online, and to the fact that a model cannot learn what it does not see.

What has the region built?

The response began with models designed around Arabic rather than translated into it:

• Jais: a 13-billion-parameter model trained on a mixture of Arabic and English. In their 2023 paper its developers say it is better than the open Arabic and multilingual models then available at Arabic knowledge and reasoning.
• ALLaM: its 2024 paper describes expanding the vocabulary and training on Arabic and English together, and its authors report that this added Arabic without "catastrophic forgetting" of English. In August 2025 HUMAIN launched a chat application in the Kingdom powered by ALLaM 34B, and said that specialist reviewers took part in aligning it with Islamic values and regional particularities. That is the company's statement; the source includes no independent evaluation of it.
• Falcon Arabic: in 2025 the Technology Innovation Institute in Abu Dhabi announced a model trained on native Arabic text covering the standard language and dialects, and described it as among the best Arabic models; the source gives no detailed benchmarks to back this.

All of these are real steps that count in their makers' favour, and the region has moved the question from "do Arabic models do well?" to "how well?". But judging which is better needs independent tests with Arabic measures, not press releases.

Fluency is not understanding

A machine may produce a grammatically sound sentence while missing a metonymy or a cultural context. Studies have documented a lean towards Western references even in Arabic and multilingual models; the details are in the article "Algorithmic Bias". Polished phrasing does not guarantee faithful meaning, and it is wise not to hand a model a religious or cultural judgement without consulting those qualified in the field.

Dialects remain an open question. Announcements speak of support for them, as in the Falcon Arabic announcement and the multi-dialect voice input in the HUMAIN app, but measuring accuracy by independent standards needs documented spoken data. Until such standards exist, the safest course is for readers to try a model in their own dialect with questions whose answers they know, then judge.

What is needed

• Good data before plentiful data: carefully spelled, content-rich texts.
• Care for diacritics and morphology: they are the basis of deep understanding, not a side detail.
• Documenting dialects: collecting spoken data with its owners' permission.
• Arabic evaluation measures: testing cultural and semantic understanding, not only the soundness of phrasing.

The question of sovereignty and the capacity to produce is wider than language, and the article "Arabic Artificial Intelligence" has addressed it. Here it is enough to say: the model that speaks Arabic has begun to arrive; the one that understands it as its people do is still being built.