The Arabic Alignment Paradox
Speak Arabic. Watch them drift apart.
When Arabic prompting reverses measured cultural alignment — a bilingual, data-driven investigation.
In this six-model study, prompting in Arabic moved all three Arabic-centric models farther from a Saudi reference profile, while all three general-purpose models moved closer.
The finding: fluency is not culture. General-purpose models scored 81.4 in Arabic; Arabic-centric models scored 109.5 — further from Saudi norms — challenging a core assumption behind regional AI.
By Bushra Asseri, Hend Alrasheed, and Areej Al-Wabil.
This interactive report needs JavaScript enabled for the full scrollytelling experience.