Elevating Global Communication with AI

What actually improved

Earlier systems translated in fragments and stitched the results together, which is why their output had the characteristic quality of being locally plausible and globally wrong. Neural translation replaced that with models that condition on the whole sentence, and large language models extended the context further still — across paragraphs, and across a conversation.

The practical consequence is that pronouns, agreement and terminology stay consistent over a long document, and that a system can be told what register to use. Asking for a formal reply to a government body and a plain one to a retail customer is now a matter of instruction rather than of building two systems.

Where Arabic is genuinely harder

Arabic is well represented in modern models, so headline quality looks strong. The difficulties are specific rather than general, and they concentrate exactly where Gulf organisations operate.

Dialect versus Modern Standard Arabic

Formal writing uses Modern Standard Arabic. Almost nobody speaks it. A customer messaging support writes in Gulf dialect, with vocabulary and structure that differ substantially from MSA — and training data skews toward MSA and toward the Egyptian and Levantine dialects that dominate published text and media. Khaleeji input is therefore the case most likely to degrade, and it is also the most common input a company in Muscat receives. Any evaluation done only on MSA will overstate real-world performance.

Ambiguity from unwritten short vowels

Arabic is normally written without diacritics, so a single written form can correspond to several words distinguished only by vowels that are not present. Readers resolve this from context effortlessly and models usually do too — but "usually" is doing real work in that sentence, and the failures are quiet, producing fluent text that means something other than the source did.

Mixed scripts and direction

Real messages contain Arabic sentences with English product names, Latin-script acronyms and Western numerals embedded inside them. Handling that correctly is partly a language problem and largely a text-handling one: bidirectional rendering, correct isolation of embedded runs, and not corrupting the order of a phone number inside an Arabic sentence.

Fluency is not accuracy

The failure mode that matters is not clumsy output. It is confident, natural-sounding output that says something the source did not. Older systems failed visibly, so reviewers stayed alert. Current systems fail invisibly, which makes review harder rather than less necessary — a reviewer skimming polished text finds fewer errors than one wading through awkward text.

Automatic quality scores compare output against reference translations and are useful for comparing systems in aggregate. They say very little about whether one particular sentence is safe to send to a customer, and they should not be used as though they do.

Deciding what to automate

The sensible split is by consequence of being wrong. High-volume, low-stakes content — product descriptions, internal documentation, first-line support replies about known issues — can be automated with sampling. Customer-facing commitments, anything a regulator reads, and anything with legal or financial effect should be machine-drafted and human-approved, which is faster than translating from scratch while keeping a person accountable.

Contracts, regulatory filings and medical instructions belong in a third category where certified human translation remains the right answer. Not because the machine cannot produce good Arabic, but because someone has to be answerable for the meaning.