By Arab Influence Leaders. This analysis was prepared with AI assistance and draws on the sources below. Original illustration by Arab Influence Leaders.
A model that answers in Arabic makes AI easy to try. Using it well requires additional skills: recognizing mistakes, designing a fair test, and deciding which information belongs in a task. Arabic AI offers ways to combine language knowledge, programming, evaluation, and product design. A small project with a clear question can demonstrate those abilities more convincingly than a collection of impressive-looking answers.
Turn a model announcement into research questions
On 9 December 2025, Mohamed bin Zayed University of Artificial Intelligence, Inception, and Cerebras announced Jais 2, an Arabic open-weight model with 70 billion parameters. The announcement describes capabilities involving Modern Standard Arabic, dialects, and switching between languages. These are descriptions from its developers, not independent proof of equal performance across every dialect or task.
A learner can translate those descriptions into focused questions. Does an answer change when the same question is phrased differently? Does a summary preserve names and numbers? What happens when the source material lacks the requested information? Such questions provide a manageable starting point without requiring a new large model or a complete retraining effort.
Learn to evaluate, not just prompt
The ArabicMMLU paper, published at ACL in 2024, introduced 14,575 multiple-choice questions across 40 tasks. Its Modern Standard Arabic material comes from educational examinations in countries across North Africa, the Levant, and the Gulf. The project illustrates the work involved in constructing an organized evaluation. Its published model results belong to that study and period; they should not be presented as today’s ranking of all models.
For a smaller exercise, test one task, such as extracting an organization and a date from public notices. Keep the expected answer and the reason it is correct. Reserve separate examples for final evaluation so that improving the instructions does not simply mean adapting them to familiar questions. Explain how another person could repeat the test and calculate the result.
Treat language data as a project in its own right
A proposed Arabic dataset project starts with usable source material, documented origins, and consistent labeling rules. If messages are classified as requests, complaints, or questions, explain how to handle one message containing several intentions. Ambiguous examples deserve an explicit rule rather than an unexplained decision.
A language specialist, programmer, and subject specialist can contribute different kinds of judgment. One checks meaning, another builds the evaluation process, and another verifies the subject matter. This is a suggested division of project work, not a claim about available vacancies or guaranteed career outcomes.
Keep original text alongside processed versions. That makes it possible to investigate whether cleaning changed a word or its meaning. Document duplication, distinguish translated material from text originally written in Arabic, and describe the topics and regions represented. A carefully explained small dataset can be a useful learning project without pretending to represent every Arabic-speaking audience.
Build knowledge in stages
UNESCO’s 2024 AI Competency Framework for Students organizes twelve competencies across human-centred thinking, ethics, techniques and applications, and system design. It describes progression through understanding, application, and creation. The framework is intended for school education. It can inspire a learning sequence here, but it is not a professional credential or a ready-made course for every learner.
Begin by understanding data, probability, and the limits of model outputs. Add practical programming skills for reading files and organizing results, then attempt a narrow project. Someone who is still learning to code can contribute evaluation criteria and review examples while developing the technical skills that the work requires.
Build a portfolio that explains decisions
A useful project portfolio states the problem, identifies its data sources, records the method, and presents errors alongside successful results. Include what remains untested: missing dialects, topics outside the sample, or operating costs that were not measured. These boundaries help a reader judge the work fairly.
For an academic or practical collaboration, offer a specific question and evidence that others can inspect. A comparison of two summarization approaches, an annotated specialist glossary, or a test of number preservation can provide a starting point. The educational value lies in building an explanation, supporting it with evidence, and revising it when a mistake appears.