Arabic AI Safety Index
The same request, written differently. Does the model behave differently?
A public NEUSAI index that measures commercially available AI models the way people in the Gulf actually write: in standard Arabic, Gulf and Egyptian dialects and Arabizi, then compares the result with English. Organisations choosing a model for an Arabic assistant or agent can request a private model-selection report.
Request a model-selection reportHow it works
- The method and weights are published before any real model is run. Any later change gets a new version number.
- We use NEUSAI's own accounts with each provider and check each provider's terms before publication.
- Each provider receives its results, the method and the test library at least 30 days before publication, and factual errors are corrected.
- No contract or sponsorship changes results or ranking.
What it covers
| Resistance to manipulation in chat | 35% of the index. |
|---|---|
| Resistance to agent hijacking through email, files or pages | 35% of the index. |
| Help with ordinary requests without needless refusal | 15% of the index. |
| Parity between Arabic and English | 15% of the index. |
Weights fixed in method version 1.0 before any real model run.
Write to us with a line about your system or your team. We reply from a named person within two working days.
Request a model-selection reportCovered by our client terms. Client terms