Trust, Safety & Content Integrity in Afghan Languages

A platform’s trust and safety stack — its classifiers, its moderators, its escalation paths — works in the languages it was built for. The terrorism recruitment, the abuse, the gender-based violence, the targeted harm in Pashto and Dari run underneath it: undetected, unclassified, and unreviewed, because nothing and no one in the pipeline reads the language. The content does not become less harmful for being unread. It becomes invisible.

The law does not accept invisibility as a defense. Under the DSA, very large platforms and search engines must assess and mitigate systemic risks across their services every year — illegal content, fundamental-rights harms, threats to civic and electoral integrity, gender-based violence, harms to minors — and the regulation states plainly that the assessment must take regional and linguistic aspects into account. The UK Online Safety Act imposes parallel duties under Ofcom. A risk you cannot read is a risk you cannot assess, and a risk you cannot assess is a finding waiting in an audit.

And the systems meant to close the gap are themselves an attack surface. The AI classifiers doing the moderation can be evaded, manipulated, and jailbroken — most easily in the low-resource languages they were trained on least. Securing them is its own discipline.

Ariana Nexus is the layer that reads what the stack cannot: native-speaker detection and classification across all 24 Afghan languages, classifier evaluation and security testing, and systemic-risk assessment and mitigation documented for the DSA and the Online Safety Act — speech-protective, reviewer-protected, and strictly within legal frameworks.