Arabic-First AI: Building From the Ground Up

Translating an English AI system into Arabic might seem like a reasonable solution, but this shortcut doesn’t work for two reasons. First, the Arab world is diverse; many versions of Arabic are used daily across the region, and not all are mutually intelligible. Second, translating from English carries over the assumptions, framings, and blind spots of English. An Arabic-first AI system has to be built from the ground up.

Governments across the Arab world have committed to AI through a series of national strategies currently in motion. The question is whether the systems being built will speak the language, reflect institutional realities, and leave room for local institutions to shape how they behave.

“Arabic-first is the starting point: the data, the model, the user flow, the product itself, all built for the region rather than adapted to it.”

Hisham Mohamed, PhD, Engineering Director, Global Public Sector at Scale

How Arabic-First AI Gets Built

An Arabic-first system has to decide at the outset which register and which region it’s being built for. This decision is skipped entirely when an English-language AI system is translated into a singular form of Arabic. Scale's Arabic-first approach sets the direction at the start, and dictates every layer that follows: the data it learns from, the dialects that data spans, the people who label that data, the model itself, the interface users touch, the evaluation that tells you whether any of it works, and the governance that surrounds it.

Arabic-first data pipelines: We start with Arabic at the source, in the registers and dialects that matter for the region, treating dropped diacritics and complex morphology as design requirements. We account for bilingual workflows, local terminology, and the language used in the documents, services, and decisions a system will support.

Labeling in the language: Our annotators read the language the way its users do. Labeling guidelines are written in Arabic, so reviewers work with the full texture of the language: dialect, tone, religious and cultural nuance, and the register expected in professional and government settings. Every label reflects how Arabic is used.

Where the task requires it, local subject-matter experts help define the guidelines, resolve ambiguity, and identify the failures that matter in context. Arabic-first labeling ensures that the people evaluating the system understand the language, domain, and consequences of getting it wrong.

Models tuned for where they’ll live: We tune on local terminology, official register, place names, and the conventions of government communication. Cultural and religious awareness is built into model behavior through training data and evaluation criteria sourced from the region. We measure success against regional use cases, because that is the benchmark that matters for this work.

End-to-end interface: The interface, the prompts, the search, the way the system reads a request and writes back, all of it is built in Arabic. We build the full experience for the region, from the first screen to the last.

Evaluation, past fluency: For Arabic-first systems, evaluation goes beyond fluency. It tests factuality, dialectal understanding, cultural fit, register, domain terminology, and whether users can safely act on the output. It also tests whether users understand the system's limits, can identify when it is wrong, and know when human judgment must override its output.

Governance held locally: Data is gathered and kept locally. End users also receive clarity and control over how data is accessed, retained, reused, audited, and governed over time.

Building for Adoption

“AI gets adopted when users can understand it, question it, apply it to their work, and feel that it respects their language, culture, and professional context.”

Norah Abokhodair, PhD, Head of Upskilling and Enablement at Scale

Most AI rollouts fail after the system is built and deployed, when the people inside the institution do not use it, do not trust it, or cannot fit it into how their work actually gets done. A system built outside an institution and translated into it produces outputs that people can read but can't fully judge. Trust is built when users have a system that fits the language they work in, the workflows they follow, and the accountability they answer to.

Arabic-first closes the gap through two choices built into the system from the start rather than layered on top. Evaluation is built to test whether users can catch the system when it's wrong, not only whether the model tends to be right. And labeling guidelines are written by local subject-matter experts, which improves the institution's ability to identify which errors are most costly, so those failure modes are found and corrected before deployment.

The goal is durable institutional capability: teams that can use AI, make informed decisions about where it belongs in their work, where it does not, and how it should change over time. That is the target that matters for the institutions these systems are meant to serve.

Arabic-First Upskilling

A deployed system gives an institution potential, but actual capability rests in the people who use the system daily. Unlocking this capability is its own piece of the work, and it has to happen in the language of the people who will team with the AI system. Scale established an Arabic-language AI literacy and upskilling program, Alif, to provide foundational Gen AI knowledge to native Arabic speakers.

The course is designed around how Arabic-speaking professionals learn, ask questions, evaluate risk, and apply AI. It starts from the conditions of the work. These workplaces are bilingual by default: policy is written in formal Arabic, the technical vocabulary of AI arrives in English, and people move between the two in a single task. The language is also sector-specific, and the setting is public, which brings trust requirements: decisions have to be traceable, sensitive data has to stay in controlled systems, and a human stays accountable for anything that affects a citizen.

Learners practice inside sector specific context. They work through how to evaluate outputs, how to protect sensitive data, and how to decide when human judgment must stay in the loop. Everyone leaves with a concrete proposal for their unit, and entities that want more can commission a run tailored to their own workflows. Alif is where people build it. The goal is people who can use AI and keep accountability for the decision. For a government, that is what makes the rest of the Arabic-first model hold over time.

AI Built for the Region

Arabic-first AI development takes the harder path: building the data, model, evaluations, and capabilities in Arabic itself, rather than delivering a finished translation. What translation misses is the ability to serve Arabic speakers as well as English systems serve the users they were built for. Building natively means a system can meet its users in their own dialect, their own register, and their own context, not an approximation of it.