Independent AI Testing and Assurance for Government Systems in Pashto, Dari and 22 More Afghan Languages

Ariana Nexus independently tests the AI systems government agencies use in Pashto, Dari and other Afghan languages, from speech recognition and machine translation to language-model assistants, OCR and screening tools, before and after deployment. Native-written test sets and severity-weighted human scoring produce evidence your AI governance board can act on, never a pass/fail seal.

Specimen 01Pronouns and grammatical person
ماPashto: “I,” in a past-tense sentence
ماDari: “we”
One written form, two meanings. A system that confuses the languages turns a first-person statement into a group statement.

Afghan-language AI assurance at a glance

Independent testing of government AI in Pashto, Dari and 22 more Afghan languages.

Systems tested
Speech, translation, interpreting, assistant, OCR, search and screening AI
Languages
Pashto and Dari in depth; 22 more Afghan languages by tier
Deliverables
Test report, evidence package, model-card annex, monitoring plan
When
Pre-deployment AI testing, acceptance before award, and monitoring in production
How agencies buy
Registered in SAM.gov; direct where rules allow, or through primes and BPA holders

What is Afghan-language AI assurance?

Afghan-language AI assurance is independent testing of an AI system's accuracy, safety, security and fairness in Pashto, Dari and other Afghan languages, using native-written test sets, human baselines and severity-weighted scoring, to produce evidence an agency can use to deploy, fix, restrict, monitor or stop the system.

It is independent verification and validation (IV&V) for AI, aligned to the NIST AI RMF and its TEVV practice: an AI audit for government systems that ends in evidence, not a seal.

Under Executive Order 14434 (September 2026), new federal documents say “Super Intelligence (SI)” for AI. The testing is the same.

Federal Register, October 2026
The columned Pennsylvania Avenue facade of the National Archives Building in Washington, D.C., in low warm light

What OMB M-25-21 requires for high-impact AI, and the evidence we produce

OMB M-25-21 compliance starts with a presumption: AI translation is high-impact when its output is legally binding or directly informs an agency decision. High-impact AI must meet seven minimum practices, required since April 3, 2026, or be safely discontinued. High-impact AI testing in Pashto and Dari supplies the evidence.

Requirement
What we test
Evidence
Requirement

Conduct pre-deployment testing

What we test
Accuracy, security and disparity on native-written items
Evidence
Test report
Requirement

Complete an AI impact assessment

What we test
Error modes, affected users and data quality per language
Evidence
Inputs for your agency’s independent reviewer
Requirement

Conduct ongoing monitoring for performance and potential adverse impacts

What we test
Drift, and re-tests after updates
Evidence
Monitoring plan and re-test schedule
Requirement

Ensure adequate human training and assessment

What we test
Whether staff catch AI errors in Pashto and Dari
Evidence
Error-recognition results
Requirement

Provide additional human oversight, intervention and accountability

What we test
Where an interpreter must review output, and when cases escalate
Evidence
Review thresholds and escalation map
Requirement

Offer consistent remedies or appeals

What we test
Whether an AI error can be traced and corrected on appeal
Evidence
Traceability findings
Requirement

Consult and incorporate feedback from end users and the public

What we test
Whether feedback channels work in Pashto and Dari
Evidence
Feedback-channel findings

Practice names follow M-25-21 §4(b). Its independent review stays inside the agency; we supply evidence, not that review.

DHS’s 2025 AI use case inventory lists four translation systems as high-impact AI. Office of Management and Budget

Acquisition, model cards and language access

Instrument
What we test
Evidence
Instrument

M-25-22 AI acquisition: vendor access for independent evaluation

What we test
Vendor claims and acceptance criteria
Evidence
Clause language and acceptance results
Instrument

M-26-04: model cards for LLMs

What we test
Cross-language benchmarks where the memo calls them necessary
Evidence
Model-card annex
Instrument

Justice Department language-access memo, July 2025

What we test
Whether AI translation is accurate enough to use responsibly
Evidence
Accuracy and severity findings

Documented failures of AI in Pashto and Dari

Pashto AI and Dari AI fail in documented ways. Each record names the error type our tests target.

  1. 2020

    United States
    A machine translation tool changed “I” to “we” in a Pashto-speaking asylum seeker’s statement; a U.S. court denied her claim over the mismatch.
    Error typePronouns and grammatical person
    Rest of World, April 2023
  2. April 2023

    Germany
    The federal migration office stopped applying dialect-recognition software to Pashto. It still applies it to Dari and Persian.
    Error typeDialect-identification claims
    Verfassungsblog, June 2026
  3. 2023

    Research
    English to Pashto, GPT-4 scored 26.7 chrF and NLLB 39.4 on the FLORES-200 benchmark.
    Error typePashto machine translation accuracy
    Robinson et al., WMT 2023
  4. 2025

    Research
    Leading commercial models read synthetic printed Pashto at 33% to 70% word accuracy.
    Error typeOCR of Pashto documents
    Haq, Zhang and Khan, May 2025
  5. 2026

    Research
    Small models often “over-corrected text to the Tehrani standard” when generating Dari.
    Error typeAfghan Dari, not Iranian Persian
    SilkRoadNLP workshop, 2026
  6. 2026

    Research
    No Whisper model produced Pashto-specific script in over 0.8% of test utterances. Fine-tuned models reporting about 14% word error rate scored 32.5% to 59% on new audio; the author maintains them.
    Error typeScript fidelity and vendor claims
    Rahman, arXiv, April 2026
  7. April–July 2026

    United Kingdom
    In first-stage Home Office spot checks, 45 of 203 AI asylum case summaries (22%) failed for factual errors.
    Error typeSummarizer accuracy
    Resultsense, October 2026

Pashto machine translation accuracy, English to Pashto

chrF score, higher is better
System
chrF, English to Pashto
System

NLLB

chrF, English to Pashto
39.4
System

GPT-4

chrF, English to Pashto
26.7

chrF measures character overlap with a reference, not meaning or severity.

Robinson et al., WMT 2023

Coverage gaps, as of November 2026

  • Neither Azure AI Speech nor Google Cloud Speech-to-Text lists a Dari speech locale.Tests for: Persian fallback in Dari speechMicrosoft Learn, November 2026
  • None of eleven multilingual safety and jailbreak benchmarks we reviewed includes Pashto or Dari.Tests for: Red-team coverageAriana Nexus research, November 2026

AI testing and evaluation (AI T&E) services for government

Every AI system an agency uses in Pashto and Dari, tested before deployment, at acquisition and in production.

  1. Pre-deployment testing for high-impact AI

    • Speech, translation, interpreting apps, assistants, summarizers, OCR, search and classification
    • Native-written test sets built around your mission
    • Black-box testing by querying the service, as M-25-21 allows without model access
    • A human baseline from qualified interpreters and translators
    • Severity-weighted analysis of which errors change a decision
  2. Security and safety testing: AI red teaming in Pashto and Dari

    • Jailbreaks, prompt injection, unsafe outputs and data leakage
    • Misuse scenarios specific to your mission
  3. Fairness and disparity testing (AI bias testing)

    • Gaps against English, between Pashto and Dari, and across dialects, speakers, ages, genders and channels
    • Multilingual AI evaluation where M-26-04 says cross-language benchmarks are necessary
  4. Procurement support and model card review

    • Requirements and evaluation criteria for solicitations
    • Test-and-monitor clauses under M-25-22
    • Vendor-claim verification and model-card review under M-26-04
    • Acceptance testing before award or an option
  5. Governance evidence for AI impact assessments

    • Inputs to AI impact assessments
    • Evidence packages for AI governance boards and Chief AI Officers
    • Records of human oversight, remedies, appeals and feedback
  6. Ongoing monitoring

    • Re-tests and drift checks after model or vendor updates
    • Sampled human review and incident review in production
  7. Human oversight design and staff training

    • When a qualified interpreter must review AI output
    • How staff spot AI errors and escalate cases
  8. National-security systems, through prime contractors

    • Support aligned to the NSPM-11 test methods once published
    • Personnel holding the clearances each contract requires
  9. Published method and benchmark

    • Our method, published and dated by edition
    • A native-written Pashto and Dari government-AI benchmark, with a university co-author
  10. Courts, allied agencies and state AI laws

    • Machine-generated evidence for courts, should proposed Federal Rule of Evidence 707 be adopted
    • Allied protection and asylum agencies
    • State AI laws as they take effect
Speech audio waveforms on color-coded tracks in audio editing software, seen close up on a computer screen

What we will never do

Independence is what an agency buys, so the limits are written down.

  • Never what we touched

    We never test, assure or report on a system we built, deployed, sold, trained or supplied data for. The Ethics & Safety Board checks conflicts, declared on every report.
  • No seals

    No certificates, seals, badges or passes: evidence and findings only.
  • No political work

    No political or campaign work, and no testing for parties or partisan actors.
  • No surveillance

    No surveillance or targeting of any community; we build no monitoring capability.
  • No judgments about people

    No origin, identity or credibility determinations about individuals.
  • No tainted data

    No data from channels controlled by the de facto authorities in Afghanistan. Test data is native-written, licensed or consented.
  • No overclaiming

    Classified work only through prime contractors, with cleared personnel. We hold no facility clearance and claim none.
  • Interest declared

    The firm also provides human interpreting. Every report says so; no finding recommends our services.

How we test Pashto and Dari AI: method, metrics and limits

Six stages aligned to the NIST AI RMF, each ending in a deliverable. Low-resource language AI evaluation needs human judgment, so metrics never stand alone.

  1. 01

    Scope

    System, mission, users and risk. Is it high-impact?
    Output: Test plan
  2. 02

    Build

    Linguists write mission items; the Cultural Compliance Bureau reviews them. Items stay held out.
    Output: Test-set specification
  3. 03

    Run

    Automated metrics and blind human scoring, beside a human baseline.
    Output: Scored results
  4. 04

    Analyze

    Severity, disparity and security findings, with confidence intervals.
    Output: Findings register
  5. 05

    Report

    Findings, evidence package and model-card annex.
    Output: Test report
  6. 06

    Monitor

    Re-tests, drift checks and sampled human review.
    Output: Monitoring plan

Inside a native-written test set

Every item is written in Pashto or Dari, never translated from English, and targets an error that changes meaning.

  • Specimen 01Pronouns and grammatical person
    Pashtoما ولیدلI saw them
    Dariما دیدیمWe saw
    ما is “I” in this Pashto sentence and “we” in Dari.
  • Specimen 02Kinship terms
    Pashtoد تره زوی
    Dariپسر کاکا
    Both: paternal uncle’s son
    English says “cousin” and loses the side of the family.
  • Specimen 03Names and transliteration
    Pashtoګل احمد
    Dariگل احمد
    Both: Gul Ahmad
    Pashto writes ګ and Dari writes گ: one name to match.
  • Specimen 04Military and police ranks
    Pashtoډګروال
    Dariدگروال
    Both: colonel
    Iranian Persian says سرهنگ; Afghan Dari does not.
  • Specimen 05Solar Hijri dates
    Pashto۱۳ لړم ۱۴۰۵
    Dari۱۳ عقرب ۱۴۰۵
    Both: 4 November 2026
    In Iran the same day is ۱۳ آبان.
  • Specimen 06Place names
    Pashtoکندهار
    Dariقندهار
    Both: Kandahar
    Two spellings, one city.
  • Specimen 07Code-switching
    Pashtoزما اپوینټمنټ کینسل شو
    Dariاپوینتمنت من کنسل شد
    Both: my appointment was cancelled
    English loanwords mid-sentence must survive.

What each metric sees, and what it misses

Metric
Sees
Misses
Use
Metric

Word error rate (WER)

Sees
Word-level transcription errors
Misses
Meaning
Use
Speech, with human review
Metric

chrF

Sees
Character overlap with a reference
Misses
Meaning and severity; unreliable in low-resource languages alone
Use
Trends, never alone
Metric

COMET

Sees
Learned similarity to a reference
Misses
Not validated for Pashto and Dari
Use
Only where checked by people
Metric

Severity-weighted human scoring

Sees
Whether an error changes a decision
Misses
Needs calibrated raters and time
Use
The anchor for every finding

The severity ladder

  1. Critical

    Changes a decision: “I” becomes “we.”
  2. Major

    Meaning changed; the reader could act wrongly.
  3. Minor

    Slightly off; the reader still acts correctly.
  4. Cosmetic

    Style only; meaning intact.

We report the critical-error rate by language, dialect and channel, beside a human baseline.

What you receive: AI test report, evidence package and monitoring plan

Documents your governance board, contracting officer and program office can use without us.

  • Test report

    Findings by language, dialect, channel and severity, with method, limits, human baseline and a unique report ID.
  • Evidence package

    For your AI governance board and Chief AI Officer, mapped to M-25-21.
  • Model-card annex

    Pashto and Dari results for a vendor model card under M-26-04.
  • Monitoring plan

    What to re-test, how often, and what triggers a re-test.
  • Re-test schedule

    Dated, and tied to model and vendor updates.
IllustrativeIllustrative format. Not a real system or result.

Test report, finding F-03

System
Speech-to-text, benefits contact center
Language
Dari (Afghanistan), telephone audio
Test set
400 native-written items, held out
Critical errors

3.0%

Major errors

7.5%

Word error rate

31%

Human baseline, critical

0.5%

Critical: “I” rendered as “we” in first-person statements.
Changes who an account is attributed to.
Options: interpreter review, fix and re-test, restrict, or pause.
Declared interest: the firm provides interpreting; no provider is recommended.

What the evidence does not establish

  • Safety for every use: findings hold for what was tested.
  • Future versions: updates call for a re-test.
  • Anything about individuals: we assess tools, never people.

Why Ariana Nexus for independent AI testing

No accredited scheme certifies AI for Pashto or Dari, and no public multilingual jailbreak benchmark we reviewed includes them. So we write native test sets and publish our method.

Why us

Native-written, never machine-translated

Automatically translated benchmarks fail systematically. Our items are written in Pashto or Dari for the mission and reviewed by our Cultural Compliance Bureau.

How we deliver

Our own people, one point of accountability

Linguists, ML evaluators and reviewers from the firm, with your program and tech teams, on schedules built for compressed timelines.

How we differ

We test the system as it runs

When a human in the loop is why a tool is judged not high-impact, that human is the safeguard. We test both together.

Security, independence and federal registration

Testing runs inside your environment or a boundary you approve, under an NDA signed first.

Ariana Nexus is an independent advisory and professional services firm for Afghan matters in Washington, D.C., founded in 2025, that independently tests government AI systems in Pashto, Dari and 22 more Afghan languages.

  • Independence wall

    Nothing we built, deployed, sold, trained or supplied data for.
  • Conflicts declared

    On every report, with any vendor payment.
  • Ethics & Safety Board

    Six powers over every engagement:
    1. 01 Approve
    2. 02 Approve with conditions
    3. 03 Require changes
    4. 04 Pause
    5. 05 Decline
    6. 06 Stop
    The Ethics & Safety Board
  • Your environment

    Testing runs there, or in a boundary you approve.
  • No retention

    None kept; none used for training.
  • NDA first

    Before any material changes hands.
  • U.S.-located testers

    For all U.S. work.
  • Our own people

    One point of accountability.
SAM.gov

Registered

UEI

M2UDMUDFXFL9

CAGE

1Z3A3

Ariana Nexus LLC, 1717 Pennsylvania Avenue NW, 10th Floor, Washington, D.C. 20006
(202) 771-0224
Capability statement

Who tests: Pashto and Dari linguists and ML evaluators

The firm’s own people: native Pashto and Dari linguists and ML evaluators, alumni and scholars of leading universities, not bilinguals.

Hassan Ukasha, Managing Partner, Ariana Nexus

Executive oversight

Hassan Ukasha

Managing Partner · Washington, D.C.
  • B.S.

    Cornell University
  • M.P.H.

    Cornell University
Oversees the firm’s operations and every program, including Afghan-language AI assurance. He is the executive accountable for each engagement’s independence and conflict review, its scope and its final report, and the escalation point for your AI governance board, contracting officer or prime contractor.
  • Zeba Haqbani, Senior Partner, Ariana Nexus

    Zeba Haqbani

    Senior Partner
    • B.Sc.

      The University of British Columbia
    Owns the evaluation harness, scoring pipeline and evidence packaging.
  • Hussain Ahmad, Analyst, Ariana Nexus

    Hussain Ahmad

    Analyst
    • M.Eng.

      Cornell University
    • Ph.D.

      University of Chicago
    Designs the measurement: metrics, sampling, rater agreement and confidence intervals.
  • Wasil Peroz, Principal, Ariana Nexus

    Wasil Peroz

    Principal
    • B.A.

      Milli University
    • M.Sc.

      Otto-von-Guericke University Magdeburg
    Maps findings to M-25-21, the NIST AI RMF and contract clauses.
  • Maryam Safi, Principal, Ariana Nexus

    Maryam Safi

    Principal
    • B.A.

      Cornell University
    Leads public-sector delivery and governance-board evidence packages.
  • Language-specialist bench

    Native linguists who write and score test items.
  • Cultural Compliance Bureau

    Reviews every native-script item for variety and register.
  • Cleared-linguist bench

    For prime-contractor work that requires cleared personnel.
  • Community connections

    Afghan community members, students, alumni and scholars take part with consent, so test sets carry real names, dialects and code-switching.
An empty university law library reading room with long tables, brass lamps, chandeliers and a tall Gothic window

Languages and where we work

Pashto and Dari in depth; 22 more Afghan languages at the depth available speakers and material allow, confirmed in writing first.

1In depth

Full test sets, human baseline, red teaming, monitoring.
  • Pashto پښتو
  • Dari دری

2Targeted

Mission items and human scoring.
  • Uzbeki
  • Turkmeni
  • Balochi
  • Pashayi
  • Aimaq

3By assessment

Depth set after reviewing speakers and material.
  • The other 17 languages listed below
Iranian 12
  • Pashto پښتو
  • Dari دری
  • Aimaq
  • Balochi
  • Ormuri
  • Parachi
  • Wakhi
  • Shughni
  • Sanglechi
  • Ishkashimi
  • Munji
  • Yidgha
Turkic 3
  • Uzbeki
  • Turkmeni
  • Kyrgyz
Indo-Aryan 3
  • Pashayi
  • Gawarbati
  • Tirahi
Nuristani 5
  • Nuristani (Ashkun group)
  • Kati
  • Prasun
  • Waigali
  • Tregami
Dravidian 1
  • Brahui
Dialects are a test dimension: southern, northern and central Pashto; Kabuli, Herati, Hazaragi and Badakhshani Dari.
Also English and 13 international languages:
  • French
  • German
  • Spanish
  • Italian
  • Hebrew
  • Urdu
  • Arabic
  • Hindi
  • Portuguese
  • Chinese
  • Russian
  • Turkish
  • Swedish
Federal, state and local agencies in all 50 states and D.C.:
  • Alabama
  • Alaska
  • Arizona
  • Arkansas
  • California
  • Colorado
  • Connecticut
  • Delaware
  • Florida
  • Georgia
  • Hawaii
  • Idaho
  • Illinois
  • Indiana
  • Iowa
  • Kansas
  • Kentucky
  • Louisiana
  • Maine
  • Maryland
  • Massachusetts
  • Michigan
  • Minnesota
  • Mississippi
  • Missouri
  • Montana
  • Nebraska
  • Nevada
  • New Hampshire
  • New Jersey
  • New Mexico
  • New York
  • North Carolina
  • North Dakota
  • Ohio
  • Oklahoma
  • Oregon
  • Pennsylvania
  • Rhode Island
  • South Carolina
  • South Dakota
  • Tennessee
  • Texas
  • Utah
  • Vermont
  • Virginia
  • Washington
  • West Virginia
  • Wisconsin
  • Wyoming
  • District of Columbia
Allied governments in Canada, Europe and Australia, by project.
An empty courtroom with wooden pews, a central aisle and the judge’s bench at the U.S. Custom House in Philadelphia

Questions agencies ask about AI testing in Pashto and Dari

Direct answers for agencies and primes.

What is high-impact AI under OMB M-25-21?

AI whose output is a principal basis for decisions or actions with a legal, material, binding or significant effect on rights, services, safety, critical infrastructure or strategic assets. M-25-21 presumes AI translation high-impact when its output is legally binding or directly informs an agency decision.

How do you test machine translation in Pashto and Dari?

We run native-written, mission-specific items through the system, and qualified translators score each output blind for meaning, severity and variety against a human baseline. Items probe names, kinship terms, ranks, pronouns, dates and code-switching.

Why not use machine-translated test sets?

Because they flatter the system: research finds automatically translated benchmarks fail systematically. Our items are written in Pashto and Dari by native linguists and reviewed by our Cultural Compliance Bureau.

How do you measure accuracy, and what does each metric miss?

Word error rate for speech, chrF for translation trends, COMET only where checked against human scores, and severity-weighted human scoring as the anchor. Automated metrics miss meaning, and in low-resource languages they need human checks.

Are you independent of AI vendors?

Yes. We never test a system we built, deployed, sold, trained or supplied data for. Conflicts, including any vendor payment, are declared on every report and checked by our Ethics & Safety Board.

Do you certify AI systems?

No. No accredited scheme exists for Pashto or Dari AI, so a certificate would claim more than testing can prove. We provide evidence; decisions to deploy, fix, restrict or stop stay with the agency.

What does an AI test report include?

Scope and method, test-set specification, results by language, dialect, channel and severity, a human baseline, security and disparity findings, metric limits and our declared interest, plus an evidence package, model-card annex, monitoring plan and re-test schedule.

Can you test classified or national-security systems?

Only through prime contractors, with personnel holding the clearances each contract requires. NSPM-11 directs NSA and DNI test methods for these systems; we align to them once published. We hold no facility clearance and claim none.

How do agencies buy independent AI testing?

We are registered in SAM.gov (UEI M2UDMUDFXFL9, CAGE 1Z3A3). Agencies buy directly where rules allow, or through prime contractors and BPA holders, including holders of the Chief Digital and AI Office’s independent test and evaluation BPA. We claim no vehicle we do not hold.

How fast can you work?

The method is built for compressed evaluation timelines, and critical findings are reported as soon as they are confirmed. We agree a written schedule at scoping, sized to the system, languages and risk.

How do you protect our data during testing?

Testing runs inside your environment or a boundary you approve, under an NDA signed first. We retain nothing afterward and never train on your data. Testers on U.S. work are in the United States.

Who are the testers?

Native Pashto and Dari linguists working with machine-learning evaluators, alumni and scholars of leading universities. Our Cultural Compliance Bureau reviews every native-script item. Hassan Ukasha, Managing Partner, oversees the program.

Does Azure or Google support Dari?

Partly. Azure AI Translator handles Dari text and Google Translate added Dari in 2024, but as of November 2026 neither Azure AI Speech nor Google Cloud Speech-to-Text lists a Dari speech locale. Systems that route Dari through Iranian Persian drift.

Do you test tools that claim to identify a speaker’s dialect or origin?

Yes, and we report their limits. We make no origin, identity or credibility determinations about individuals. Germany’s migration office stopped applying such software to Pashto in April 2023.

Sources

  1. OMB M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public TrustOffice of Management and Budget, April 2025. Checked 4 November 2026.
  2. Trump Administration Revamps Guidance on Federal Use and Procurement of AIWiley Rein, April 2025. Checked 4 November 2026.
  3. OMB M-25-22, Driving Efficient Acquisition of Artificial Intelligence in GovernmentOffice of Management and Budget, April 2025. Checked 4 November 2026.
  4. OMB memoranda on use and acquisition of AI: what M-25-22 means for government contractorsPilieroMazza, 2025. Checked 4 November 2026.
  5. OMB M-26-04, Increasing Public Trust in Artificial Intelligence Through Unbiased AI PrinciplesOffice of Management and Budget, December 2025. Checked 4 November 2026.
  6. White House imposes transparency requirements on contractors (M-26-04)Crowell & Moring, December 2025. Checked 4 November 2026.
  7. National Security Presidential Memorandum 11 (NSPM-11)GovInfo, June 2026. Checked 4 November 2026.
  8. Independent test and evaluation BPA, one-page overviewChief Digital and AI Office. Checked 4 November 2026.
  9. DHS AI use case inventory: USCISU.S. Department of Homeland Security. Checked 4 November 2026.
  10. 2025 Federal Agency AI Use Case Inventory (consolidated dataset)Office of Management and Budget. Checked 4 November 2026.
  11. Attorney General memorandum implementing Executive Order 14224U.S. Department of Justice, July 2025. Checked 4 November 2026.
  12. Executive Order 14434Federal Register, October 2026. Checked 4 November 2026.
  13. DHS and ICE determinations on high-impact AIAmerican Immigration Council. Checked 4 November 2026.
  14. AI translation errors and Afghan asylum claimsRest of World, April 2023. Checked 4 November 2026.
  15. Dialect recognition software (DIAS) and the lawVerfassungsblog, June 2026. Checked 4 November 2026.
  16. ChatGPT MT: Competitive for High- (but Not Low-) Resource LanguagesRobinson et al., WMT 2023. Checked 4 November 2026.
  17. PsOCR: Benchmarking Large Multimodal Models for OCR in PashtoHaq, Zhang and Khan, May 2025. Checked 4 November 2026.
  18. One Language, Three of Its Voices: Evaluating Multilingual LLMs Across Persian, Dari, and TajikiSilkRoadNLP workshop, 2026. Checked 4 November 2026.
  19. Benchmarking Multilingual Speech Models on PashtoRahman, arXiv, April 2026. Checked 4 November 2026.
  20. Home Office asylum AI summaries fail checksResultsense, October 2026. Checked 4 November 2026.
  21. Translation as a Scalable Proxy for Multilingual EvaluationarXiv, January 2026. Checked 4 November 2026.
  22. Language and voice support for Azure AI SpeechMicrosoft Learn, November 2026. Checked 4 November 2026.
  23. Speech-to-Text supported languagesGoogle Cloud, 2026. Checked 4 November 2026.
  24. AI Risk Management FrameworkNational Institute of Standards and Technology. Checked 4 November 2026.
  25. Review of eleven multilingual safety and jailbreak benchmarksAriana Nexus research, November 2026.

Facts on this page were checked on 4 November 2026.

Start with the system and the mission

Tell us what the system does, who it serves and which languages it meets. We will tell you what can be tested and what the evidence will prove.

Discuss an evaluation