Arabic-first AI, not Arabic-translated AI
Why models trained-then-translated fail Arab institutions — and what to demand instead.
- 01Translation is not nativity — insist on native corpora
- 02Ask for dialect-specific evals, not just MSA
- 03Ten side-by-side outputs beat any deck
Most "Arabic AI" today is English AI wearing a keffiyeh. The model reasons in English, then translates. The result reads like a diplomatic cable written by someone who has never eaten mansaf.
Arabic-first means the corpus, the evaluation set, the reviewers, and the failure cases are Arabic from day zero. Dialect matters. Register matters. Whether the model knows the difference between fus'ha for a khutbah and fus'ha for a press release matters.
When you procure an Arabic AI system, ask three questions. What percentage of pre-training was native Arabic, not back-translated? Who wrote the eval set — native speakers or annotators paid by the hour? Show me ten outputs in Emirati dialect and ten in MSA, side by side.
If the vendor cannot answer, they built a translator, not a model. Pay accordingly.
Voice guardrails for AI content
How to encode taste as a rubric so agents don't drift into slop.
Context is the product
Prompts are cheap; the context you feed the model is the moat.
The zag rule
If your competitor could run your prompt and get the same output, you don't have a brand.