Ariana Nexus trains the people who build AI products and run online platforms to read Afghan content correctly: trust and safety policy, content moderation, product, machine learning and legal teams. The training covers Pashto, Dari and 22 more Afghan languages, Islam as it is practiced in Afghanistan, ethnicity, gender, four decades of war, and the Afghan diaspora. It is live, built on your own policies and cases, and taught by Afghan scholars.
One Afghan post. Seven kinds of context decide how to read it.
د مصنوعي ځیرکتیا او د باور او خوندیتوب ټیمونو لپاره د افغاني کلتور او دین د شالید روزنه
آموزش زمینهٔ فرهنگی و دینی افغانستان برای تیمهای هوش مصنوعی و اعتماد و ایمنی
Afghan cultural and religious context training is professional training for the teams that write content policy, moderate content, design products and train AI models. It teaches how Afghans speak and write, how they practice Islam, how they name themselves and count time, how ethnic and sectarian lines work, and how forty years of war shape everyday language. The goal is correct decisions about Afghan content, Afghan users and AI answers about Afghanistan.
It is not general diversity training and it is not a language course. It is decision training. Every module ends in the question a moderator, a policy writer, a product manager or an annotator actually faces: leave it up or take it down, escalate or close, ship or fix, label it safe or unsafe.
Afghan content is decided by context. A model or a moderator without that context gets the decision wrong in both directions: it removes grief, poetry and news, and it misses threats, hate and fraud.
These are the nine places where Afghan context decides the outcome. Each one produces errors in both directions: too much enforcement against ordinary Afghans, and too little against real harm.
Failure point 01
Allahu Akbar, inshallah, mashallah, shaheed and Quranic verses appear in condolences, weddings, exam results and news. Keyword-led systems and untrained reviewers treat them as signals of terrorism.
A trained team reads genre and register first, and knows which added signals turn devotion or grief into praise or incitement.
Connected to
Failure point 02
Reporting, commentary and even criticism of the Taliban is removed under dangerous organizations policies, while real propaganda in Pashto poetry and taranas passes because nobody recognized the genre.
A trained team applies the news reporting, neutral discussion and condemnation allowances correctly, and recognizes the forms propaganda actually takes.
Failure point 03
Slurs and coded insults aimed at Hazaras, Pashtuns, Tajiks, Uzbeks, Shia Muslims, Sikhs and Hindus do not survive machine translation. They arrive in English as harmless words.
A trained team knows the terms, the history behind them, and the difference between in-group use, reclaimed use and attack.
Failure point 04
A photo of a woman without a headscarf, a leaked phone number or a rumor about a girl can lead to honor-based violence. Severity scales built on Western norms rate these as low harm.
A trained team calibrates severity for Afghan women, girls and LGBTQ Afghans, and sends cases with offline risk to escalation.
Failure point 05
Posts that name former interpreters, soldiers, judges, journalists or aid workers who “worked with the foreigners” are life-safety risks that look like plain sentences.
A trained team recognizes doxxing of at-risk Afghans and knows which groups have been most exposed since August 2021.
Connected to
Failure point 06
Pashto is routed to Urdu or Arabic reviewers. Dari is treated as Iranian Persian. Hazaragi, Uzbeki and Balochi are not recognized at all. Pashto and Dari typed in Latin letters defeat language detection.
A trained team identifies the language, the variety and the script before anything else, and knows when to ask for a native reviewer.
Failure point 07
Many Afghans have one name and no family name. U.S. documents show FNU, first name unknown. Tazkiras often record a year or an estimated age, so very large numbers of Afghans share a 1 January birthday. Afghanistan counts years in the Solar Hijri calendar.
A trained team designs sign-up, search, age checks, account recovery and identity verification that work for Afghan users.
Connected to
Failure point 08
Models give confident answers on contested facts such as ethnic population shares, refuse ordinary questions about Islam, flatten Hanafi and Jafari practice into one answer, and miss how Afghans describe distress, war trauma and moral injury.
A trained team writes better guidelines, test prompts, refusal rules and escalation paths for Afghan users, Afghan patients and veterans of the Afghan war.
Connected to
Failure point 09
Ramadan, Ashura, Nowruz, 15 August and the days after an attack or an earthquake change what people post and what it means. Queues spike and error rates rise.
A trained team plans staffing, guidance and product behavior around the Afghan calendar.
Each module runs 90 minutes and ends in decisions, not definitions. Modules are selected by role.
Module 01. Part 1, language, script and identity
Pashto, Dari and 22 more languages in five families: where each is spoken, how to tell Pashto from Urdu and Arabic, Dari from Iranian Persian and Tajik, and why Hazaragi is a variety of Dari. Dialects, Latin-letter typing and mixed-language posts.
Connected to
Module 02. Part 1, language, script and identity
Right-to-left text, the letters only Pashto uses, Persian and Arabic numerals, the Solar Hijri calendar, single names, fathers’ names, honorifics, tazkiras, FNU and 1 January birthdays. What this means for forms, search, age checks and identity verification.
Connected to
Module 03. Part 1, language, script and identity
Poetry, proverbs, landays, humor, insults and irony. Formal and informal address. What Afghan users post on each kind of platform, and what machine translation does to it.
Connected to
Module 04. Part 2, religion
The Sunni Hanafi majority, Twelver and Ismaili Shia communities, Sufi orders and shrines, Afghan Sikhs and Hindus, and who holds religious authority. What is devotional, what is political and what is militant.
Connected to
Module 05. Part 2, religion
Shaheed, jihad, kafir, takfir, Allahu Akbar, Quranic quotation, taranas and nasheeds. How to separate devotion, grief, commentary and news from praise, threat and incitement. Why an accusation of blasphemy or apostasy is itself a threat.
Connected to
Module 06. Part 2, religion
Ramadan, Eid al-Fitr, Eid al-Adha, Muharram and Ashura, Mawlid, Nowruz, Independence Day and 15 August. Planning staffing, guidance and product behavior around them.
Connected to
Module 07. Part 3, society, war and harm
Pashtun, Tajik, Hazara, Uzbek, Turkmen, Baloch, Aimaq, Nuristani, Pashayi and other communities. How slurs, dehumanizing language and coded insults work in Pashto and Dari, why population figures are contested, and how violence against Hazaras shapes what is posted.
Connected to
Module 08. Part 3, society, war and harm
Namus, mahram, purdah, women’s images and voices, forced and child marriage, and the restrictions on women and girls since 2021. Severity calibration for intimate images, doxxing, outing of LGBTQ Afghans and harassment.
Connected to
Module 09. Part 3, society, war and harm
From 1978 to today: the Soviet war, the civil war, the first Taliban rule, the Republic and August 2021. The vocabulary, flags and symbols of each side. The Taliban as de facto authorities, Islamic State Khorasan Province and the armed opposition. News reporting and condemnation exceptions, commemoration versus glorification, and sanctions basics.
Connected to
Module 10. Part 4, people and products
Afghans in the United States, Germany, Canada, the United Kingdom, Australia, Iran and Pakistan. Special Immigrant Visas, humanitarian parole, deportations and family separation. Former interpreters, soldiers, judges, journalists and activists. Doxxing, threats across borders and the scams that target evacuees.
Connected to
Module 11. Part 4, people and products
How models answer questions about Islam, Afghan history, ethnicity and women’s rights. Contested facts, refusal and over-refusal, tone and forms of address. Health and mental health answers for Afghan patients: idioms of distress such as jigar khun and asabi, war trauma, moral injury and safe messaging. Image generation and stereotypes.
Connected to
Module 12. Part 4, people and products
Calibration exercises on your anonymized queue samples, model outputs and labeling guidelines. We log every place your written policy is silent or ambiguous for Afghan content and hand over recommended language.
Connected to
Core for this teamRecommended
In September 2022 the Oversight Board overturned Meta’s removal of a Facebook post by an Urdu-language newspaper that reported a Taliban announcement about girls’ schools. The Board found the post was news reporting, which the policy allows, and noted that fewer than 50 Urdu-speaking reviewers were assigned to the system meant to catch such errors.
Oversight Board, case 2022-005-FB-UA
In March 2024 the Oversight Board advised Meta to end its blanket removal of the word shaheed when it refers to designated individuals. The Board found that the word has several meanings, that it accounted for more removals than any other word or phrase, and that removal should depend on other signals of violence. Afghans use the word every day in Pashto and Dari.
Oversight Board, policy advisory opinion 2023-01
Research from the Center for Democracy and Technology found that multilingual language models are trained mostly on English text and carry English-language assumptions into other languages, which limits how well they moderate content in languages such as Pashto and Dari.
Center for Democracy and Technology, Lost in Translation, 2023
This is the kind of sentence the training is built on. It is a composite written for teaching, not taken from any user or client.
خدای دې وبښي، شهید شو. انا لله و انا الیه راجعون.
خدا بیامرزد، شهید شد. انا لله و انا الیه راجعون.
May God forgive him. He was martyred. We belong to God and to Him we return.
What an automated system sees: the word martyr, an Arabic religious formula and a death. Likely label: praise of a dangerous organization.
The sentence is Pashto or Dari. The last clause is Arabic, a verse of the Quran (2:156) that Muslims everywhere say when they hear of any death.
This is a condolence. It is what people write under the photo of a relative, a neighbor or a stranger who has died.
In Afghan usage shaheed is given to civilians killed in bombings, to people who die in accidents and earthquakes, and to soldiers on every side. It honors the dead. It does not endorse a cause.
After more than forty years of war, most Afghan families have someone they call shaheed. The word is on street signs, schools and memorial days.
No violation. What would change it: the name of a designated person together with praise of the act, a call to follow it, or other signals of violence.
شهید
Martyr, also witness. An honorific for the dead, used for civilians, accident victims and fighters on every side.
Most removals of this word are errors.
Taught in
جهاد
Struggle. In Afghan history it names the 1980s war against the Soviet army. Mujahid and mujahideen come from it.
Historical and devotional use far outnumbers militant use.
Taught in
الله اکبر
God is greatest. Said in prayer, in joy, in shock and in grief.
Not a signal of violence on its own.
Taught in
انشاءالله، ماشاءالله
If God wills, and what God has willed. Everyday speech about the future and about good news.
Classifiers trained on extremist text give them too much weight.
Taught in
ناموس
Family honor, tied to the women of the family.
Explains why an image or a rumor can carry a risk of violence.
Taught in
محرم
A close male relative: father, brother, husband or son. Current rules in Afghanistan require women to travel with one.
Appears in travel, aid, health and education content.
Taught in
تذکره
The Afghan national identity document, on paper or electronic.
Often has no exact date of birth and no family name.
Taught in
First name unknown. Printed on U.S. documents for Afghans who have a single name.
Breaks sign-up forms, name matching and account recovery.
Taught in
هجری شمسی
Afghanistan’s civil calendar since 1922. The year starts at Nowruz, on or about 21 March, and month names differ in Pashto and Dari. Since 2022 the de facto authorities have used the lunar Hijri calendar in official business.
Dates need conversion, and the calendar itself is contested.
Taught in
نوروز
New Year and the first day of spring. Widely celebrated, and removed as a public holiday by the de facto authorities in 2022.
A celebration for most Afghans and a point of dispute for some.
Taught in
عاشورا
The tenth day of Muharram. A day of mourning for Shia Muslims, and a repeated target of attacks on Hazara and Shia communities.
Graphic and sectarian content rises around it.
Taught in
ترانه
A chanted song without instruments. Used for poetry and national songs, and by the Taliban for propaganda.
The genre is not the violation. The content is.
Taught in
Policy writers and market specialists who own the rules on dangerous organizations, hate speech, violent speech, harassment and privacy, and need those rules to work in Pashto and Dari.
Moderators, quality analysts, trainers and outsourced vendor teams who make the decisions at volume, including reviewers who see Afghan content only through machine translation.
Product managers, designers, localization leads and conversation designers who ship assistants, search, sign-up and identity flows to Afghan users.
Annotation leads, preference-data and safety-classifier teams, and evaluators who write labeling guidelines for Afghan-language data.
Counsel and policy leads who answer to regulators on systemic risk, AI literacy, sanctions and human rights due diligence.
The people on call when Afghanistan is in the news: an attack, an earthquake, a deportation wave, an anniversary.
Fees are fixed per engagement and quoted after a scoping conversation. They depend on the format, the number of cohorts and how much we build from your own material. Every engagement is delivered to one organization. We do not sell seats in open courses.
We learn your products, your Afghan user base, your policies and your team structure. You learn what we would cover and what we would not.
Before we build anything, we read your community standards or usage policies, enforcement guidelines, labeling instructions and a sample of anonymized decisions or model outputs.
Participants work through short Afghan-content scenarios. Results are reported by team and used to decide where the time goes.
We select modules by role and write cases from your own surfaces. Every example, translation and glossary entry is reviewed by our Cultural Compliance Bureau before it reaches your team.
Afghan scholars teach every session live. Sessions are built around decisions: participants decide, compare, and hear how an Afghan reader sees the same content. No graphic media is shown.
A final assessment, a completion record for each participant, the policy gap log, reference cards, glossary and sensitive-dates calendar. Office hours stay open for 90 days.
We agree the measures with you before the first session and report them at the end. We do not publish client results, and we do not promise a number before we have seen your baseline.
Share of Afghan-content scenarios decided correctly, before and after, by team.
Agreement between reviewers on a calibration set of Pashto and Dari items, before and after.
How often devotional, condolence and news content is wrongly actioned in the calibration set.
How often ethnic hate, doxxing of people at risk and gender-based threats are missed in the calibration set.
Appeal overturn rates, escalation volumes and time to decision for Afghan-language content, tracked by you over the following quarter.
Training supports these obligations. It does not by itself establish compliance, and we work alongside your counsel. References are stated as of September 2026.
This program is taught by the partners and principals of Ariana Nexus, not by contract trainers. They come from the Afghan community and studied at Cornell, Brown, the University of Chicago, the University of British Columbia and other leading universities. Between them they hold graduate degrees in public health, psychology, law and engineering, and they work every day on Afghan-language AI evaluation, interpreting quality and compliance.
That combination is the point. A bilingual employee can tell you what a sentence says. A scholar who is also a native speaker can tell you what it does: who it is aimed at, what it will lead to, and how a Pashtun reader in Kandahar, a Hazara reader in Kabul and an Afghan family in Virginia would each take it.
Program oversight
Managing Partner
B.S., Cornell University. M.P.H., Cornell University.
Hassan Ukasha oversees the firm’s operations and this program. He sets the standard every session is taught to, approves the scope of each engagement, and reviews the policy language a client’s team leaves with. He is based in Washington, D.C.
Most cultural training is written by generalists and delivered from slides. Most Afghan expertise available to a platform is one bilingual reviewer. This program is neither.
The faculty come from the Afghan community and studied at Cornell, Brown, the University of Chicago, the University of British Columbia and other leading universities, in public health, psychology, law, and AI and data engineering. We do not send a bilingual. We send a scholar who is also a native speaker.
We read your policies before we write a slide. Your team practices on your own surfaces and cases, under NDA, and leaves with recommended policy language.
When a case needs a Hazaragi, Uzbeki or Balochi reader, we bring that reader from our own bench. Nothing is brokered to a third party.
Our Cultural Compliance Bureau reviews each example, translation and glossary entry for accuracy and balance before it is taught.
Afghan politics is divided by ethnicity, sect and the war. The program teaches how Pashtun, Tajik, Hazara, Uzbek and other communities, Sunni and Shia, read the same content, and says plainly where Afghans disagree.
The team that teaches this program also evaluates language models, runs red teaming and tests AI interpreting products in Pashto and Dari. What we find in testing goes into what we teach.
Every session is designed and delivered by our own people. We have no office or operations in Afghanistan, and nothing is routed through channels controlled by the de facto authorities.
It is specific to Afghan content, Afghan users and Afghan markets.
Participants do not learn Pashto or Dari. They learn when they need someone who reads it.
We teach the context around sanctions, designation and regulation. Your counsel decides what your obligations are.
Training raises the floor. It does not replace Pashto and Dari reviewers, evaluation or red teaming.
We hand over a glossary with context. We do not supply blocklists, because lists without context are how most of these errors start.
No accredited certification exists for this subject. We issue completion records and say so plainly.
The program teaches Pashto and Dari in depth, because they carry most Afghan content online. The other 22 languages are taught at orientation depth: what they are, who speaks them, how to recognize them and when to bring in a native reader.
Pashto, Dari, Hazaragi (a variety of Dari), Aimaq, Balochi, Ormuri, Parachi, Wakhi, Shughni, Sanglechi, Ishkashimi, Munji, Yidgha
Uzbeki, Turkmeni, Kyrgyz
Pashayi, Gawarbati, Tirahi
Nuristani (Ashkun group), Kati, Prasun, Waigali
Brahui
It is live professional training that teaches policy, content moderation, product, machine learning and legal teams how to read Afghan content correctly. It covers Pashto, Dari and 22 more Afghan languages, Islam as it is practiced in Afghanistan, ethnicity, gender, four decades of war and the Afghan diaspora, and every module ends in the decisions those teams actually make.
Afghanistan is not an Arab country and is not in the Middle East. Its main languages, Pashto and Dari, are Iranian languages, not Arabic. Most Afghans follow the Hanafi school of Sunni Islam, with large Shia communities, and the poetry, names, calendar and forty years of war are specific to the country. Training built on Arabic-speaking countries gets Afghan content wrong in predictable ways.
There are four formats: a 90-minute executive briefing, a one-day team intensive, a six-week full program of 12 modules, and a continuing program of quarterly refreshers. All are live and instructor-led, delivered remotely or on site. Instruction is in English, with Pashto and Dari material.
Yes. Before we build anything we read your policies, enforcement guidelines and labeling instructions under NDA, and we write the cases from your own product surfaces and anonymized decisions. The last module is a workshop on your own cases, and you receive a log of every place your written policy is silent or ambiguous for Afghan content.
Yes. Vendor moderators, quality analysts, trainers and annotation teams can join your cohorts or have their own, under your NDA. We also run sessions for your own trainers so the material reaches new hires.
It supports them. Article 4 of the EU AI Act, as amended in July 2026, asks providers and deployers to take measures that support staff AI literacy in the context their systems are used in. The Digital Services Act and Ofcom’s codes under the Online Safety Act expect trained, adequately resourced moderation that accounts for language. We provide role-based training and completion records. Training alone does not establish compliance, and your counsel decides what your obligations are.
Pashto and Dari in depth, and 22 more Afghan languages at orientation depth, including Hazaragi as a variety of Dari, Uzbeki, Turkmeni, Balochi, Pashayi and the Nuristani and Pamir languages. It covers Pashtun, Tajik, Hazara, Uzbek, Turkmen, Baloch and other communities, Sunni and Shia Muslims, Afghan Sikhs and Hindus, and the Afghan diaspora in North America, Europe, Australia, Iran and Pakistan.
We teach recognition and decision-making inside your policy. Participants learn how the Taliban, Islamic State Khorasan Province and other armed groups communicate, how to apply news reporting, discussion and condemnation allowances, and how to tell commemoration from glorification. We do not show graphic media, and sanctions questions go to your counsel.
Both. One module is given to AI assistants, search and generative products: how models answer questions about Islam, Afghan history, ethnicity and women’s rights, where they over-refuse, and how they handle health and mental health questions from Afghan patients, including war trauma and moral injury among Afghans and veterans of the Afghan war.
Evaluation and red teaming test the model. This training teaches the people who write the policies, label the data and make the decisions. Most clients need both, and the same Ariana Nexus team delivers them, so what we find in testing goes into what we teach.
No graphic images, audio or video are shown. Cases are described in text, content notes are given in advance, and any participant can step out without explanation. Sessions are planned with Diana Ayubi, Psy.D., who sets and reviews the participant wellbeing protocol. Many participants are Afghans or veterans themselves, and the program is written with that in mind.
Fees are fixed per engagement and quoted after a scoping conversation. They depend on the format, the number of cohorts and how much material we build from your own policies and cases. We do not sell individual seats or open-enrollment courses.
Ariana Nexus, a consulting and professional services firm in Washington, D.C. The program is taught by the firm’s own partners and principals, Afghan alumni of Cornell, Brown, the University of Chicago and the University of British Columbia, and every example is reviewed by the firm’s Cultural Compliance Bureau. The office is at 1717 Pennsylvania Avenue NW, 10th Floor, Washington, D.C. 20006.
Tell us which products, teams and policies are in scope. We will tell you what the training would cover, what it would not, and which format fits. Scoping runs under NDA, and we meet in person at our Washington, D.C. office where that helps.
1717 Pennsylvania Avenue NW, 10th Floor, Washington, D.C. 20006
(202) 771-0224
Ariana Nexus trains and advises. It does not certify individuals or platforms, and it does not give legal advice. Regulatory references on this page are stated as of September 2026. Last reviewed September 2026.
