Before a machine-translated discharge instruction, portal message or chatbot reply reaches an Afghan patient, a qualified human translator from Ariana Nexus reviews it, scores every error by severity, corrects it and signs the record.
For health programs covered by Section 1557, federal rules already require this review for critical patient text. We make it fast, measurable and specific to Afghan Pashto, Dari and every language Afghan patients speak.
Translation engines learn from text that already exists online, and little of it is Afghan. Most Persian text online is Iranian, much of the Pashto online comes from Pakistan and is not medical, and several Afghan languages are rarely written at all. The output still reads fluently, which is the danger: a patient cannot tell a fluent error from a correct instruction.
Any patient-facing text in an Afghan language that a machine or an AI model produced, reviewed against its English source and in the form the patient will actually see.
Every item leaves a record your auditors can read, tied to the reviewer who signed it.
Two constructed examples of the errors we see most often. Pashto first, then Dari.
Six steps, the same for every item, so a result in Pashto can be compared with a result in Dari or Uzbeki.
Section 1557 of the Affordable Care Act and its 2024 rule draw the line for AI translation in health care. When a covered entity uses machine translation for text that is critical to a patient’s rights, benefits or meaningful access, where accuracy is essential, or where the source is complex, non-literal or technical, a qualified human translator must review the translation. Discharge instructions, medication directions and consent forms meet at least one of those conditions.
No U.S. certification exam tests translation from English into Pashto or Dari, so buyers are left to take a vendor’s word. Ariana Nexus answers with a written reviewer standard, reviewers who meet it and a signed record for every item.
The program is overseen by Hassan Ukasha, Managing Partner, and delivered by partners and managers educated at Cornell, Brown, the University of Chicago and the University of British Columbia. The reviewers they lead are Afghan college graduates, not bilingual volunteers.
Oversees the firm’s operations and the AI translation review program, including engagement governance, the reviewer standard and data protection. Grew up in Herat.
We review written output in all 24 languages and say plainly when machine translation is not a sound option. Several of these languages are rarely written, so the right answer may be a human translation or a recorded message rather than an engine. For Hazaragi speakers, we review Afghan Dari for Hazaragi readers or record the message in Hazaragi.
Regions are indicative and used for matching; they are not a census. Names are shown in Arabic script as commonly written; several smaller languages have no standardized spelling. Engine coverage changes often. We check the specific tool you use at intake.
Everyday health words, as Afghan patients read them and as engines often return them.
Not on its own. In a 2021 study of emergency department instructions, Google Translate was 94% accurate for Spanish but 67.5% for Persian (Farsi), the closest published measure for Dari. Afghan Dari differs from Iranian Persian in everyday health words, so the output needs review by a qualified human translator before a patient relies on it.
It allows the tool, not unreviewed output. Under 45 CFR 92.201(c)(3), if a covered entity uses machine translation for text that is critical to a patient’s rights, benefits or meaningful access, where accuracy is essential, or where the language is complex, non-literal or technical, a qualified human translator must review the translation.
The Section 1557 rule, at 45 CFR 92.4, describes a qualified translator as someone who follows translator ethics, including confidentiality, has shown proficiency in written English and the other language, and translates accurately and impartially with the specialized terms the text needs. Our reviewers are assessed in the exact variety they review and trained in medical terminology.
You can use it to draft. We treat translation by AI models as machine translation, because it is automated, text-based translation, so critical patient text still needs review by a qualified human translator. AI models also write fluent text that hides errors and can change meaning between runs, so each output needs its own review.
They are closely related varieties of Persian, but not the same in daily use. Afghan Dari uses different words for many health terms, such as شفاخانه for hospital and داکتر for doctor, where Iranian Persian uses بیمارستان and پزشک. Tools trained mostly on Iranian text tend to produce the Iranian forms.
Post-editing (MTPE) corrects machine output until it reads well. A quality review also measures it: every error is categorized and scored by severity, and the result is signed and recorded. You learn how good the tool is, not only what the corrected text says.
We follow ISO 5060:2024, the international guidance for evaluating translation output, and the MQM error typology. Each error is typed, such as accuracy, terminology or locale, and weighted as critical, major or minor. Critical errors block release until they are corrected and checked again.
Pashto is widely offered. Dari is offered by some tools but often returned as Iranian Persian. Uzbeki, Turkmeni and Kyrgyz are usually produced in another script, while Afghan communities write them in Arabic script. Most other Afghan languages have no reliable engine, and several are rarely written; for those we recommend human translation or recorded audio.
Yes. We match the language, the exact variety, the city or region your patients come from, and the reviewer’s gender where the subject calls for it. Clinical content goes to reviewers with degrees close to medicine.
A business associate agreement is signed before any patient text is shared. Reviewers work only in approved systems, never paste patient text into public AI tools, and handle only the text the review needs.
Executive Order 14224, signed in March 2025, designates English as the official language but does not require agencies or health programs to stop offering services in other languages. Section 1557 and its language access rule, including the machine translation review requirement, remain in effect.
AB 3030 requires California health facilities, clinics and physician offices that use generative AI for communications about a patient’s clinical information to include a disclaimer and instructions for reaching a person, unless a licensed or certified health care provider reads and reviews the message. Our language review does not replace that clinical review, but it makes sure the message and the disclaimer are right in Pashto or Dari.
No U.S. certification exam tests translation from English into Pashto or Dari. That is why we publish our own reviewer standard: a degree, an assessment in the exact variety, medical training, ethics and HIPAA training, supervised entry and periodic review.
Scope and handling are agreed under NDA, with a business associate agreement in place before any patient text is shared.
