Ariana Nexus provides Pashto and Dari content moderation and trust and safety policy localization for social platforms, messaging apps, marketplaces and AI labs. We rewrite your community guidelines and enforcement rules for Afghan context, build the lexicons, review high-risk and ambiguous content, calibrate your moderators, and test your classifiers and LLM moderation against the same rules.
Pashto and Dari in Perso-Arabic script and Romanized form, mixed with Urdu, Iranian Persian and English, in text, images, audio and video — from Afghanistan, Pakistan and the Afghan diaspora.
Most platforms moderate Pashto and Dari with rules written in English, classifiers trained on other languages and reviewers hired for fluency rather than judgment. The result fails in both directions: harmful content stays up, and news, documentation and criticism come down.
Six service lines, built into one program around your own policies.
A gold top rule marks harms that are critical by default. Final tiers are set with you.
Maps to dangerous organizations; terrorism
Maps to hate speech
Maps to harassment; intimate image abuse
Maps to violence and incitement
Maps to privacy; doxxing
Maps to violent and graphic content
Maps to fraud; misinformation
Maps to human exploitation
Maps to child sexual exploitation
Maps to regulated goods
Maps to suicide and self-injury
Maps to platform manipulation
Pashto on the right, Dari on the left, wherever they sit side by side.
Eleven letters of the Pashto alphabet that standard Persian and Arabic keyboards do not carry, with their Unicode code points. Text typed without them arrives in variant spellings.
| Pair | Letters and code points | Where it comes from |
|---|---|---|
| PairPashto gaf and Persian gaf | Letters and code pointsګU+06ABandگU+06AF | Where it comes fromPashto typed on a Persian keyboard |
| PairPashto ṭe and Urdu ṭe, two encodings | Letters and code pointsټU+067CandٹU+0679 | Where it comes fromPashto typed on an Urdu keyboard in Pakistan |
| PairPersian ye and Arabic ye | Letters and code pointsیU+06CCandيU+064A | Where it comes fromKeyboard variants in Dari; different letters in Pashto |
| PairPersian kaf and Arabic kaf | Letters and code pointsکU+06A9andكU+0643 | Where it comes fromArabic keyboards and copied text |
| PairTwo forms of the digit four | Letters and code points۴U+06F4and٤U+0664 | Where it comes fromDigit variants slip phone numbers past filters |
Normalization has to know which language it is reading: in Dari, ی and ي are keyboard variants of one letter; in Pashto they are different letters and must not be merged.
One word, four common Latin spellings. Southern speakers say Pashto; northern speakers say Pakhto.
The zero-width non-joiner (U+200C) belongs inside many Dari words. It is invisible, and it is also used to split a banned term so a filter no longer matches it.
په پښتو او دري ژبو کې د منځپانګې څارنه
نظارت بر محتوا به زبانهای پشتو و دری
Content moderation in Pashto and Dari
Output: Baseline error sample and gap log
Output: Localized enforcement guidelines
Output: Versioned lexicon and golden sets
Output: Calibration record
Output: Decision records and monthly quality report
Output: Monthly report and change logs
Reported monthly by harm area, language, script and dialect.
correct / reviewed
κ · α
overturned / appealed
TP / (TP+FP) · TP / (TP+FN)
t(action) − t(receipt)
Δ terms / month
| Tier | What it covers | Response target |
|---|---|---|
| TierT1 Critical | What it coversCredible threats, terrorist content, child safety, intimate image abuse, doxxing of people at risk | Response targetAction within 1 hour of receipt |
| TierT2 High | What it coversHate speech, graphic violence, scams, smuggling, regulated goods | Response targetAction within 4 hours |
| TierT3 Standard | What it coversRoutine policy review and appeals | Response targetAction within 1 business day |
EU TCO removal order
Our tier-2 target
TAKE IT DOWN Act request
DSA risk assessment cycle
Logarithmic scale, one hour to one year. Gold marks our tier-2 target.
| Regulation | Jurisdiction | What it requires | Where Pashto and Dari come in | Status |
|---|---|---|---|---|
| RegulationDigital Services Act7 | JurisdictionEuropean Union | What it requiresNotice and action, statements of reasons and internal appeals; for very large platforms, annual systemic risk assessments that account for regional and linguistic aspects (Article 34) and moderator reporting by official EU language (Article 42). | Where Pashto and Dari come inPashto and Dari risk inputs, localized statements of reasons and appeal texts, and capacity evidence for languages Article 42 does not break out. | StatusAll platforms since 17 February 2024; very large platforms since August 2023 |
| RegulationTerrorist Content Online Regulation (EU) 2021/7848 | JurisdictionEuropean Union | What it requiresRemove or disable terrorist content within one hour of a removal order. | Where Pashto and Dari come inPashto and Dari identification of designated-group propaganda and one-hour escalation paths. | StatusApplies since 7 June 2022 |
| RegulationOnline Safety Act 20239 | JurisdictionUnited Kingdom | What it requiresIllegal harms risk assessments and duties to remove illegal content quickly; children’s safety duties. | Where Pashto and Dari come inPashto and Dari inputs to illegal harms risk assessments; review of terrorism, threats and intimate image abuse. | StatusIllegal content duties since 17 March 2025 |
| RegulationTAKE IT DOWN Act (Pub. L. 119-12)10 | JurisdictionUnited States | What it requiresRemove non-consensual intimate images, including AI-generated ones, within 48 hours of a valid request. | Where Pashto and Dari come inPashto and Dari request handling and victim-aware review. | StatusFTC enforcement since 19 May 2026 |
| RegulationOnline Safety Act 202111 | JurisdictionAustralia | What it requiresBasic Online Safety Expectations and removal notices from the eSafety Commissioner. | Where Pashto and Dari come inPashto and Dari review for Australia’s Afghan communities. | StatusIn force since January 2022 |
We work alongside your counsel and do not give legal advice. Regulatory dates on this page are reviewed every quarter.
Illustrative engagement types, not client case studies.
The people accountable for this work are scholars and alumni of Cornell University, the University of Chicago, the University of British Columbia and Otto-von-Guericke University Magdeburg who come from the Afghan community. They write the policy your moderators apply, measure the results, and answer for every critical decision.
Managing Partner, Ariana Nexus
Languages: Pashto and Dari (native); English, Urdu and Hindi; Arabic (working).
Senior Partner
Principal
Principal
Principal
| Person | Owns on your program | Signs off |
|---|---|---|
| PersonHassan Ukasha | Owns on your programScope, review standard and executive escalation | Signs offEvery localized policy before release |
| PersonZeba Haqbani | Owns on your programSecure review environment, tool integration and data flow | Signs offEnvironment access and the deletion record at close |
| PersonHussain Ahmad | Owns on your programGolden sets, sampling plans, and classifier and LLM moderation evaluation | Signs offEach evaluation report |
| PersonWasil Peroz | Owns on your programRegulatory mapping and designated-entity rules | Signs offDangerous-organization guidance before release |
| PersonMaryam Safi | Owns on your programPashto and Dari lexicons, worked examples and the gap log | Signs offEach monthly lexicon release |
Diana Ayubi, Psy.D., Engagement Manager, sets and reviews the reviewer wellbeing protocol: exposure limits, rotation off graphic queues, and scheduled debriefs. Reviewers who come from the communities in Afghan war footage face a higher risk of trauma and moral injury, and the protocol is built for that risk.
We wrote the review standard, we train our reviewers and your vendors’ moderators to it, and every client sees it.
| Criterion | Machine translation and generic classifiers | Outsourced bilingual seats | Ariana Nexus |
|---|---|---|---|
| Pashto and Dari policy | Machine translation and generic classifiersWord-for-word translation of English rules | Outsourced bilingual seatsThe vendor’s generic guidelines | Ariana NexusEnforcement rules rewritten for Afghan context, with examples and edge cases |
| Coded language and script variants | Machine translation and generic classifiersMissed unless spelled as trained | Outsourced bilingual seatsDepends on the individual moderator | Ariana NexusVersioned lexicon with Romanized and keyboard variants, updated monthly |
| Taliban-related news and criticism | Machine translation and generic classifiersOver-removal is common | Outsourced bilingual seatsInconsistent | Ariana NexusDesignated-entity guidance with tested news and condemnation rules |
| What gets measured | Machine translation and generic classifiersAggregate accuracy | Outsourced bilingual seatsHandle time and volume | Ariana NexusPrecision, recall and agreement by harm, language, script and dialect |
| Regulatory evidence | Machine translation and generic classifiersNone | Outsourced bilingual seatsLimited | Ariana NexusInputs for DSA, TCO, Online Safety Act and TAKE IT DOWN duties |
| Reviewer wellbeing | Machine translation and generic classifiersNot applicable | Outsourced bilingual seatsVaries by vendor | Ariana NexusClinical protocol with exposure limits and debriefs |
It is the review of user-generated content written or spoken in Pashto and Dari against a platform’s rules: removing, labeling, restricting or keeping posts, comments, images, audio and video. Done well, it uses rules localized for Afghan context, lexicons that cover script and dialect variants, and reviewers who read the languages as first languages.
Policy localization adapts a platform’s community guidelines and internal enforcement rules to a language and culture. It goes beyond translation: definitions, examples, edge cases and exceptions are rewritten so a Pashto or Dari reviewer, or a model, reaches the decision the policy intends.
Not on its own today. Automated classifiers and large language models perform best in English and lose accuracy in low-resource languages, especially with dialect, slang, Romanized text and audio. We measure your models against a localized policy and design the human review that covers what they miss.
Dari and Iranian Persian (Farsi) are closely related but not interchangeable for moderation. Vocabulary, slang, political references and spelling conventions differ, and a reviewer trained on Iranian content will misread Afghan context. We moderate Afghanistan Dari, including Hazaragi, a variety of Dari, and read Farsi where it is mixed into Dari content.
We apply your dangerous organizations or designated-entity policy with Afghan-context guidance that separates praise, support and representation from news reporting, neutral discussion and condemnation. Two Oversight Board cases involving posts about the Taliban and girls’ education show how often the second group is removed by mistake.
Yes. Our Pashto content moderation covers Pashto in Perso-Arabic script and in Latin letters, Southern and Northern varieties, and usage in Khyber Pakhtunkhwa, Balochistan and Karachi as well as Afghanistan, including code-switching with Urdu and English.
We staff expert review for high-severity, ambiguous and appealed Pashto and Dari content, and we calibrate the in-house or outsourced moderators who handle volume. We do not sell per-item seats; engagements are scoped after a confidential review of your policies and queues.
The EU Digital Services Act, including risk assessments that must account for linguistic aspects; the EU Terrorist Content Online Regulation’s one-hour removal duty; the UK Online Safety Act’s illegal harms duties; the U.S. TAKE IT DOWN Act’s 48-hour removal rule; and Australia’s Online Safety Act. We work alongside your counsel and do not give legal advice.
Our protocol, set by a clinical psychologist, limits exposure to graphic content, rotates reviewers off the hardest queues and schedules debriefs. Many reviewers come from the communities shown in Afghan war and violence footage, which raises the risk of trauma and moral injury; the protocol is designed for that risk.
We sign your NDA before we see anything, work inside your moderation tools or a secure environment you control, keep no copies of user content after review, provide a written deletion record at close, and never use client content to train models.
Content in Uzbeki, Turkmeni, Balochi, Pashayi and the other Afghan languages we work in can be read on request, with response times set by language. This service is scoped to Pashto and Dari, where platform volume and risk are highest.
Ariana Nexus is headquartered in Washington, D.C. Engagements start with a confidential briefing and a review of your policies and a sample of your Pashto and Dari queues. We sign your NDA first.
We recruit reviewers directly and train them to our standard. Reviewers looking for work can apply through Ariana Nexus careers.
Terms as we use them in policies, lexicons and reports.
We sign your NDA before we see a single policy or queue item.
Request a confidential briefing