KurdishAIin2026:whatworksandwhatdoesn't

Short version: AI can help with Kurdish text today, but it is clearly weaker than in English or Arabic — and weakest in Badini, the Kurdish we speak in Duhok. Listening to Kurdish and reading it from photos are weaker still. Here is what works today, what doesn't, and what I tell clients. Every tool claim below was checked against the vendor's own language list in October 2026.

Writing and chat: usable, with a human check

Claude, ChatGPT and Gemini can draft everyday Kurdish business text: replies to customers, product descriptions, summaries. In my own testing on Kurdish tasks, Claude does best, then ChatGPT, then Gemini. But quality is well behind English and Arabic, and Badini is the weakest, because far less written Badini exists for models to learn from.

Independent research says the same thing: good, not perfect. A study in the UHD Journal of Science and Technology tested ChatGPT with Kurdish multiple-choice questions and found about 70% accuracy on average, with the weakest results on history.

What to do: let AI draft, and have a Kurdish speaker check anything customers will read. For a real system, give the model your own business words and test it on real customer messages, not demo examples. My setup for Kurdish chatbots is in Building multilingual AI chatbots.

Translation: works, with a Badini catch

Google Translate has two Kurdish options: Kurmanji and Sorani. Badini falls between them. It is a Kurmanji dialect, but in Duhok we write it in Arabic letters, while Google's Kurmanji option uses Latin letters. Choose Sorani instead and you get a different dialect.

What to do: for Sorani and Latin-script Kurmanji, machine translation is a good first draft. For Badini, translate, then have a Badini speaker fix the wording.

Speech-to-text: Sorani only

This is the biggest gap. OpenAI's Whisper, the most widely used open speech model, has no Kurdish at all. Google Cloud Speech-to-Text lists Central Kurdish (Sorani), but not Kurmanji or Badini.

What to do: if your customers send Badini voice notes on WhatsApp, no tool I checked turns them into reliable text today. Ask for typed messages for orders and bookings, or keep a person on voice notes. Do not buy a system that promises Badini speech without a live demo on your own recordings.

Reading photos and scans (OCR): test first

OCR is the technology that reads text from a photo or a scanned page — invoices, IDs, receipts.

• Tesseract, the most widely used free OCR engine, has a Kurmanji model for Latin letters only. It has no official model for Kurdish in Arabic letters.
• Microsoft Azure AI Vision lists Kurdish in Arabic letters, Kurdish in Latin letters, and Sorani.

What to do: test 20 of your real documents before you pay for anything. Clean printed text has a chance. Handwriting is much harder.

What this means for your business

• Kurdish replies, posts and product text: yes, as drafts, always with a human check.
• Translating documents: yes for Sorani and Latin-script Kurmanji. For Badini, add a human editor.
• Badini voice notes and calls: not ready yet.
• Scanning Kurdish paperwork: maybe. Test on your own documents first.

Why Badini is behind

AI learns from text and recordings on the internet, and there is very little Badini online. Every website, dataset and recording in Badini helps the next models learn it. That is one reason this site is written in Kurdish as well as English and Arabic.

Remember: AI can draft Kurdish, with a human check. Listening and reading photos are not ready for Badini. Test on your own messages before you buy.

Sources

• OpenAI Whisper — supported languages
• Google Cloud Speech-to-Text — supported languages
• Google Cloud Translation — supported languages
• Tesseract OCR — trained language models
• Microsoft Azure AI Vision — language support
• UHD Journal of Science and Technology — ChatGPT and Kurdish study