Will the Machine Preserve Our Eloquence? AI and the Arabic Language

Ask a language model for an Arabic essay, and it comes back with sound grammar and no spelling slips in its hamzas. Is that enough? The difference between correct speech and eloquent speech is the difference between a map and a city. In that difference it is decided whether the machine will preserve our eloquence or dissolve it in the mould of another language.

A slight presence on the web

W3Techs, in a survey dated 22 October 2026, reports that English is the content language of about 49.5% of websites whose language is known, and Arabic of about 0.6%. This is a figure for websites, not for model training data, so it cannot be applied directly, but it indicates a general direction. The authors of a 2020 study found a clear gap between languages in available resources and questioned the claim that models are “language agnostic”. (See the article “Arabic AI” for sovereignty and data.)

The Arabic word on the scales

Models learn by splitting text into small units called tokens. If the splitter was designed around other languages, texts equal in meaning need different numbers of tokens. Petrov and colleagues (2023) measured this and found differences of up to 15 times for some languages, even with multilingual tokenizers. Three consequences follow: higher cost, since pricing is tied to tokens; longer processing time; and a context that holds fewer words. The study does not concern Arabic alone, so we should not attribute a figure to Arabic in particular, but languages with rich derivation and dense inflection are likely to be affected more, and that deserves measurement specific to Arabic.

Models built from within

This does not mean fate is sealed. Jais (2023) was announced, a 13-billion-parameter model trained on Arabic and English together, with far less English data than comparable English models use. A research team published the ALLaM paper (2024), which says it expanded the model's vocabulary, combined Arabic and English in training, used parallel translated data, and reached advanced results on Arabic benchmarks such as MMLU Arabic. These are results reported by their authors, and they need independent evaluation and new metrics that match real use.

Correctness is not eloquence

Most tests measure what can be counted: a correct answer, a sound sentence. Arabic is a language of exposition, suggestion, metonymy and double meaning, and its beauty lies in a measured departure from the familiar, while the model by its statistical nature leans towards the most probable familiar. It may thus master correctness and fall short in eloquence; this is a logical inference, not an experimental result. The danger is that the young grow used to a flattened language and take it for the original.

What should we do?

First, build high-quality Arabic corpora from the best of heritage and sober contemporary writing, respecting authors' rights. Second, support models designed for Arabic's features, and ask for metrics that test style, not correctness alone. Third, use the tools for vocalisation, proofreading and digitising manuscripts, which are useful if their output is carefully reviewed. Fourth, cultivate in students the habit of authentic reading and of not accepting every machine output as given, for the machine imitates what we offer it and cannot replace the cultivation of taste.

The machine does not defeat Arabic, but its people's neglect might. It learns from what we give it, and what it returns to us is a mirror of what we write. The safest course is to sow what we wish to reap: an eloquence that is read and written, and we ask God to preserve the language of the Quran and help us serve it.