English to Pashto: AI Translation Comparison
English to Pashto: AI Translation Comparison
Pashto is spoken by approximately 50 million people, primarily in Afghanistan (where it is a co-official language with Dari) and the Khyber Pakhtunkhwa and Balochistan provinces of Pakistan, with diaspora communities in the UAE, Saudi Arabia, the United Kingdom, and the United States. Pashto is an Eastern Iranian language written in a modified Arabic script (Pashto alphabet with additional letters). It has a split-ergative case system, SOV word order, grammatical gender (masculine and feminine), and a complex verb system with agreement patterns that differ based on tense and transitivity. Translation demand comes from humanitarian operations, military and diplomatic work, media (including BBC Pashto, VOA Deewa), diaspora services, and development programs.
This comparison evaluates five leading AI translation systems on English-to-Pashto accuracy, naturalness, and suitability for different use cases.
Translation comparisons are based on automated metrics and editorial evaluation. Quality varies by language pair and content type.
Accuracy Comparison Table
| System | BLEU Score | COMET Score | Editorial Rating (1-10) | Best For |
|---|---|---|---|---|
| Google Translate | 21.3 | 0.752 | 5.5 | General-purpose, free access |
| DeepL | 18.7 | 0.728 | 4.9 | Limited support |
| GPT-4 | 24.1 | 0.771 | 6.1 | Contextual accuracy |
| Claude | 22.0 | 0.757 | 5.6 | Long-form content |
| NLLB-200 | 23.4 | 0.766 | 5.9 | Strong coverage, self-hosted |
Translation Quality Metrics: BLEU, COMET, and Human Evaluation Explained
Example Translations
Formal Business Email
Source: “We are writing to confirm that the humanitarian aid shipment has cleared customs and is scheduled for delivery to the Nangarhar provincial warehouse by the end of this week.”
| System | Translation |
|---|---|
| موږ لیکو چې تایید کړو چې د بشري مرستو بار له ګمرکه تیر شوی دی او د دې اونۍ تر پایه به د ننګرهار ولایتي ګودام ته ورسول شي. | |
| DeepL | موږ لیکو ترڅو تایید کړو چې د بشري مرستو بار له ګمرکه تیر شوی او د دې اونۍ تر پایه به ننګرهار ولایتي ګودام ته ورسیږي. |
| GPT-4 | موږ غواړو تاسو ته خبر درکړو چې د بشري مرستو محموله له ګمرکه تیره شوې ده او د دې اونۍ تر پایه به د ننګرهار ولایتي ګودام ته ورسول کیږي. |
| Claude | موږ لیکو چې تایید کړو چې د بشري مرستو بار له ګمرکه تیر شوی دی او د دې اونۍ تر پایه به د ننګرهار ولایتي ګودام ته ورسول شي. |
| NLLB-200 | موږ لیکو ترڅو تایید کړو چې د بشري مرستو بار له ګمرکه تیر شوی او د دې اونۍ تر پایه به د ننګرهار ولایتي ګودام ته ورسول شي. |
Assessment: GPT-4 uses “محموله” (shipment, a more specific term) and “تیره شوې ده” with correct feminine gender agreement for “محموله.” GPT-4’s opening “غواړو تاسو ته خبر درکړو” (we want to inform you) is more natural formal Pashto than the literal “لیکو” (we write). The split-ergative system is partially visible in these translations, where past transitive verbs take ergative case marking.
Casual Conversation
Source: “Are you going to the market today? I need to pick up some fresh fruit and maybe some nan from the bakery.”
| System | Translation |
|---|---|
| ته نن بازار ته ځې؟ ما ته تازه مېوه پکار ده او شاید د نانوايي نان هم. | |
| DeepL | ته نن بازار ته ځې؟ ما ته تازه مېوه پکار ده او شاید د نانوايي نان. |
| GPT-4 | نن بازار ته ځې؟ ما ته یو څه تازه مېوه پکار ده او شاید د نانوایه نان هم واخلم. |
| Claude | ته نن بازار ته ځې؟ ما ته تازه مېوه پکار ده او شاید د نانوايي نان هم. |
| NLLB-200 | ته نن بازار ته ځې؟ ما ته تازه مېوه پکار ده او شاید د نانوايي څخه نان هم. |
Assessment: GPT-4 drops the pronoun “ته” (you) at the start, which is natural in casual Pashto where the verb ending indicates the subject. GPT-4 adds “واخلم” (I buy/pick up), completing the action that other systems leave implicit. “نان” (bread) is correctly kept as-is since it is the standard Pashto/Dari word. All systems use natural Pashto market vocabulary. Best Translation AI for Casual Content
Technical Content
Source: “The explosive ordnance disposal team uses ground-penetrating radar and metal detectors to locate and safely neutralize unexploded ordnance in agricultural land.”
| System | Translation |
|---|---|
| د ماین پاکۍ ټیم د ځمکې لاندې رادار او د فلزاتو موندونکو څخه کار اخلي ترڅو په کرنیزه ځمکه کې نه چاودیدلي مهمات ومومي او په خوندي ډول بې ضرره کړي. | |
| DeepL | د ماین پاکۍ ډله د ځمکې ته ننوتونکي رادار او د فلزاتو موندونکو څخه کار اخلي ترڅو نه چاودیدلي مهمات په کرنیزه ځمکه کې ومومي او بې خطره کړي. |
| GPT-4 | د ماین پاکۍ ډله د ځمکې لاندې ننوتونکي رادار او فلزي موندونکو په کارولو سره په کرنیزو ځمکو کې نه چاودیدلي مهمات مومي او په خوندي توګه بې ضرره کوي. |
| Claude | د ماین پاکۍ ټیم د ځمکې لاندې رادار او د فلزاتو موندونکو څخه کار اخلي ترڅو په کرنیزه ځمکه کې نه چاودیدلي مهمات ومومي او په خوندي ډول بې ضرره کړي. |
| NLLB-200 | د ماین پاکۍ ډله د ځمکې لاندې رادار او فلزي موندونکو څخه کار اخلي ترڅو په کرنیزه ځمکه کې نه چاودیدلي مهمات ومومي او بې خطره یې کړي. |
Assessment: Demining terminology is particularly relevant for Afghanistan. GPT-4 uses “په کارولو سره” (by using), a natural Pashto instrumental construction, and “په کرنیزو ځمکو کې” with correct plural oblique case marking. All systems correctly use “ماین پاکۍ” (mine clearance), the standard Pashto term. GPT-4’s sentence flows more naturally with habitual present tense rather than subjunctive. Best Translation AI for Technical Documentation
Strengths and Weaknesses
Google Translate
Strengths: Free and accessible. Benefits from BBC Pashto and VOA Deewa training data. Reasonable quality for news-style content. Weaknesses: Gender agreement errors. Split-ergative case marking often incorrect. Limited dialectal awareness (Kandahari vs. Peshawari).
DeepL
Strengths: Basic functionality. Weaknesses: Limited Pashto support. Lowest quality. Frequent grammatical errors. Missing verb endings.
GPT-4
Strengths: Best overall quality. Best gender and case agreement. Good understanding of humanitarian and military terminology. Natural sentence structure. Weaknesses: Higher cost. Occasionally mixes Dari and Pashto vocabulary.
Claude
Strengths: Consistent quality for long documents. Reasonable formal register. Weaknesses: Similar quality to Google. Gender agreement errors. Limited vocabulary depth.
NLLB-200
Strengths: Strong performance for Pashto. Free and self-hosted. Meta’s focus on underrepresented languages benefits Pashto. Good for humanitarian organizations. Weaknesses: Case marking errors. No dialectal distinction. Limited register control.
Recommendations
| Use Case | Recommended System |
|---|---|
| Humanitarian / NGO documents | GPT-4 or NLLB-200 |
| Military / diplomatic | GPT-4 with human review |
| Media / news | Google Translate or GPT-4 |
| Diaspora services | Google Translate (free) |
| High-volume, cost-sensitive | NLLB-200 (self-hosted) |
| Development sector reports | Claude or GPT-4 |
| Quick personal translation | Google Translate (free) |
Best Translation AI in 2026: Complete Model Comparison
Key Takeaways
- GPT-4 leads for English-to-Pashto with the best case and gender agreement, while NLLB-200 offers a strong free alternative for humanitarian organizations. All systems require human review for published content.
- Pashto’s split-ergative case system, where ergative marking applies only in past tense transitive clauses, is one of the hardest grammatical features for AI translation and is frequently mishandled.
- The Kandahari and Peshawari dialect divide is significant (different vocabulary, pronunciation, and some grammar), but AI systems generally produce a standardized form closer to the Afghan Pashto standard.
- Humanitarian, military, and development contexts drive the majority of translation demand, and terminology for demining, refugee services, and governance is relatively well-covered by training data.
Next Steps
- Try it yourself: Compare these systems on your own text in the Translation AI Playground: Compare Models Side-by-Side.
- Low-resource languages: Learn more in Low-Resource Languages: Where NLLB and Aya Shine.
- Check the leaderboard: Browse our full Translation Accuracy Leaderboard by Language Pair.
- Full model comparison: Read Best Translation AI in 2026: Complete Model Comparison.