BATOOL GHAITH (ABU DHABI)

Artificial intelligence is becoming more capable in Arabic, but understanding how Emiratis actually speak — including dialect, code-switching, local expressions and cultural references — remains a more complex challenge than simply training a model on Arabic text.

Eng Ahmed Al Ameri, NEP-AI Expert for Arabic-centric LLMs and Localisation, said the gap between Arabic and English AI partly reflects the much larger volume of English-language digital content and training data developed over decades.

“The challenge is not simply that there is less Arabic in AI, it is that the Arabic data available to models does not always capture the full diversity of how Arabic is actually spoken and written,” Al Ameri told Aletihad.

Arabic itself adds another layer of complexity because it is not used in one uniform form, he said.

Models need to understand morphology, word patterns and how meaning changes depending on context. Arabic text also ranges from Modern Standard Arabic to dialectal writing and informal digital communication, Al Ameri explained.

“The next stage of Arabic AI development is not simply about translating existing AI capabilities into Arabic, it is about building models that are Arabic-centric from the ground up, models that understand the richness of the language, its dialects, its cultural context and the way Arabic speakers actually communicate.”

Understanding Emirati Arabic

Al Ameri said there has been significant progress in Arabic-language AI, but a clear gap remains between understanding Modern Standard Arabic and understanding it as it is spoken in specific local contexts.

“Emirati Arabic has its own vocabulary, expressions, idioms and cultural references,” he said. Some expressions may carry different meanings in other Arabic dialects, while others may not have direct equivalents in the modern standard.

For AI systems, accurate understanding requires exposure to authentic, high-quality Emirati language and the cultural context behind it, according to Al Ameri.

“In the UAE, it is common for people to move naturally between Emirati Arabic, Modern Standard Arabic and English — sometimes within the same conversation or even the same sentence,” he added, noting that this is especially common in professional, technical and digital environments.

An AI system, therefore, needs to understand not only each language independently, but also why a speaker moves between them and what that switch communicates, Al Ameri explained.

“A literal translation may be grammatically correct but completely miss the intended meaning, tone or cultural reference,” he said.

“The real measure of localisation is, therefore not whether an AI model can translate an Emirati expression, but whether it can understand what the speaker actually means and respond naturally in the appropriate context,” Al Ameri added.


Building UAE-specific Data

Al Ameri said UAE-specific datasets have an important role to play in improving how AI understands Emirati Arabic and local cultural context.

However, he said local datasets should complement rather than replace wider Arabic-language data.

“A strong Arabic AI ecosystem requires broad and diverse Arabic data, complemented by high-quality local datasets that capture the linguistic and cultural characteristics of individual communities,” Al Ameri said.

He also emphasised that quality matters as much as volume, as datasets need to be accurate, representative and responsibly collected, with attention to privacy, consent, intellectual property and data governance.

Al Ameri also stressed the importance of evaluating models separately from training them.

“We need reliable benchmarks that can tell us whether an AI system genuinely understands Emirati Arabic, rather than simply producing Arabic that sounds grammatically correct.”

For the UAE, developing Arabic-focused AI is also linked to building national capability, Al Ameri said the objective is both to serve Arabic speakers more naturally and to ensure the country has the expertise needed to shape strategically important technologies.

“For the UAE, sovereignty in AI does not mean developing everything independently or operating in isolation; it means having the capabilities, talent, infrastructure data and intellectual expertise to understand, develop and shape technologies that are strategically important to the country,” he said.

He pointed to Falcon-H1-Arabic, which was purpose-built for Arabic-language understanding and trained on native Arabic datasets, and Falcon-Emirati, developed specifically to understand local vocabulary, expressions, idioms and cultural context.

“Language is not simply a communication tool. It carries culture, history, identity and knowledge,” Al Ameri added.

He also indicated that AI could also help strengthen Arabic usage among younger generations if the experience feels natural and relevant.

A student could use AI to learn or research in Arabic, while young content creators could develop ideas in Arabic and explore Emirati history, literature, cultural expressions and local heritage through more personalised experiences, he noted.

“The quality of the experience matters. If Arabic AI systems provide responses that feel unnatural, overly formal or disconnected from the way young people actually communicate, they are less likely to use them,” Al Ameri said.

He also stressed that Arabic should not be treated only as a language for cultural or traditional content, but also for science, technology, entrepreneurship, research, creativity and innovation.

“When AI understands not only the language, but also the context, culture and identity behind it, technology becomes a tool for strengthening that identity rather than diluting it,” Al Ameri said.