Arabic text in AI images: why it breaks and how to fix it
Why AI image models still get Arabic letters wrong, what the tools' own notes say in 2026, and the workflow I use to add correct Arabic type to AI images.
Published 9 min read
Most AI image generators still write Arabic badly. Ask for a short Arabic word and you may get letters that don't join, dots in the wrong place, a word that runs backwards, or shapes that only look Arabic from a distance. Some 2026 models handle short text better than before, but as of October 2026, none of the official pages I read promises reliable Arabic.
The fix is not a magic prompt. It is a workflow: generate the image without text, then set the Arabic yourself as real type or real lettering in Photoshop, Illustrator, Figma or Canva. The model makes the picture; the Arabic stays in your hands.
Below I explain why the script is hard for image models, what the main tools say about it, the four approaches that work, and the step-by-step workflow I'd use, drawing on years of drawing Arabic lettering for brands. This article is part of my guide to AI in graphic design.
Why image models struggle with Arabic
If you don't read Arabic, it helps to know what makes the script different. If you do, this explains why the mistakes look the way they do.
Letters join and change shape
Arabic is written joined up, and most letters change shape depending on where they sit in a word: alone, at the start, in the middle or at the end. The letter ʿayn, for example, has four quite different forms: ع عـ ـعـ ـع. Six letters (ا د ذ ر ز و) never join to the letter after them, so a word breaks into pieces at specific places, and only there. One wrong join or one wrong form and the word stops being a word.
Dots carry the meaning
Several letters share the same basic shape and differ only in their dots. ب ت ث differ only in the number and place of dots, and at the start or middle of a word ن and ي share that shape too. ج ح خ differ by one dot. To a reader, a missing or extra dot changes the word. To a model that learned from pixels, dots look like texture: small marks it can move, multiply or drop.
Right to left, with numbers left to right
Arabic runs from right to left, but numbers and Latin words inside it run from left to right. Design software handles this with bidirectional text rules. A model has no such rules, so letters can come out mirrored, in the wrong order, or with directions mixed.
Pixels, not characters
When you type Arabic in Photoshop, a shaping engine picks the correct form of each letter and joins them. Adobe says Photoshop's Unified Text Engine uses the open-source HarfBuzz shaping engine, plus FriBidi for right-to-left and mixed text. An image model has nothing like that. It doesn't type letters; it paints pixels that resemble letters it has seen. Latin survives this better, because each letter has one shape and stands on its own.
Less Arabic to learn from
Companies rarely publish exactly what their models learned from, but their own notes point the same way. Midjourney says text works best with the standard Latin alphabet. Ideogram lists English and other Latin-script languages. It is reasonable to expect that clean Arabic text is a small part of what these models have seen. Arabic's rich calligraphic tradition probably doesn't help either: the same letter looks very different in Naskh, Kufi, Thuluth or Diwani, so the "average" letter a model learns is likely to be blurred.
What the tools say, as of October 2026
These notes come from each company's own pages. They will age, so check again before you rely on them.
- Midjourney: put text in double quotes; it works best with the Latin alphabet and short phrases.
- Ideogram: text rendering supports English and other Latin-script languages, and support for other scripts is "expanding".
- Google Nano Banana 2: can generate and translate text in many languages, but may struggle with grammar, spelling and cultural nuance. Google's own chart for the model shows that most models struggle with long text in most non-Latin scripts, and Arabic is one of the scripts on that chart.
- OpenAI ChatGPT Images 2.0 (April 2026): improved multilingual text rendering, while OpenAI's developer documentation says the model can still struggle with precise text placement and clarity.
In practice: a short Arabic word may sometimes come out right, longer Arabic usually doesn't, and nothing official promises it. Treat every Arabic letter a model writes as a draft to check. For what each tool is good at beyond text, see my guide to choosing an AI image generator for designers.
Four approaches that work
1. Try a text-focused model, for very short words
For one or two words, such as a café name on a cup in a moodboard, a model known for text can be worth a try. Give it the exact string in quotes, keep it short, and expect to check and fix it. Never use this for a final logo, a label or a price.
2. Generate the image without text, then set the type
This is the route I trust most. Ask the model for the scene with no text at all and a planned empty area where the headline will sit. Then set the Arabic as live type in a tool that handles right-to-left text properly. Photoshop and Illustrator have full Arabic support, Figma supports right-to-left and bidirectional text with a text direction control, and Canva keeps layouts intact between left-to-right and right-to-left languages.
The hardest case is text that belongs to the scene itself: a shop sign at night, a name printed on a box, a menu on a table. Here the type has to look photographed, not pasted on. Generate the scene with a blank sign or a blank pack, then place the real Arabic artwork on it in Photoshop as a smart object, with perspective, the scene's light and shadow, and a touch of its grain. It takes longer than a prompt, but it is the difference between a mockup that sells and one that embarrasses the client in front of Arabic readers.
3. Use the image as a background
For posters, social posts and presentation slides, treat the AI image as a background plate. Add a gradient or a solid area for contrast, and keep the Arabic in its own layer on top. The type stays sharp and editable, and the image can be swapped without touching the words.
4. Redraw it as vector lettering
Sometimes a model suggests an interesting composition for a word: a curve, a stack, a texture. Use it as a sketch, nothing more. Redraw the letters by hand in Illustrator with the correct forms, joins and dots, so the idea survives and the spelling is right.
A practical workflow
- Write the final Arabic first. Proofread it, decide on diacritics, and agree it with the client before any image exists.
- Decide where the text lives. Live type (social posts, ads), drawn lettering (logos, product names), or text inside the scene (a shop sign, a package).
- Prompt for space, not letters. Describe the scene, say "no text, no letters, no signage", and say where the empty area should be. I explain how I structure these prompts in prompt writing for designers.
- Clean any fake letters. Models often add letter-like marks anyway. Remove them with Generative Fill or careful retouching.
- Set the Arabic as real type. Use a proper Arabic font, check it contains every character you need, and give Arabic more line height than Latin.
- Fit text into the scene. For a sign or a package, place the type as a smart object, apply perspective, and match the light, grain and blur of the image.
- Use drawn lettering for names. If the text is a brand name, place the real vector lettering, never a model's version of it.
- Check it with a native reader. Read it letter by letter: joins, dots, the lam-alef (لا), direction, digits.
- Label AI images when you show the work, so people know what was generated.
From my work: why I draw the Arabic myself
Much of my identity work starts with an Arabic name drawn by hand, and these are exactly the cases where a model can't be trusted with the letters.
- Kababgy Al-Sultan: the name is drawn as one round seal, with the grill and its flame as its crown, and the letters of "Al-Sultan" curl like smoke. Some takeaway and menu scenes on the project page were made with AI around the real designs, and each is labelled as an AI scene. The scene can be generated; the lettering is the brand.
- Abdeen Coffee: a hand-drawn Arabic mark built like a seal. Every step was drawn by hand before it touched the computer.
- Fanash: a snack pack with the Arabic name lettered by hand with a bounce, so it reads as playful to a child and clear to a parent.
In all three, the character of the brand lives in the exact shape of each letter. A model that drops one dot or breaks one join doesn't produce a rough version of the logo. It produces a different word.
A checklist before you publish
| Check | Why it matters |
|---|---|
| Every dot is present and in place | One dot changes the letter, and the word |
| Letters join where they should, and only there | Wrong joins make the word unreadable |
| Each letter uses the right form for its position | Start, middle, end and alone forms differ |
| Lam-alef (لا) appears as one ligature | It is required in Arabic, not optional |
| The text runs right to left | Mirrored or reversed words are common errors |
| Numbers and Latin words run left to right | Mixed direction needs real bidirectional handling |
| A native reader has read it aloud | The fastest way to catch what you missed |
Conclusion
AI image models are useful for scenes, light and atmosphere, but in October 2026 they are still not reliable writers of Arabic. Don't fight the model for correct letters. Let it make the picture, leave a clear space, and set the Arabic yourself as real type or hand-drawn lettering. Your client gets an image that reads correctly, and text that can still be edited tomorrow.
If you need Arabic lettering or a bilingual identity that holds up in every medium, AI scenes included, start a project here.
Questions people ask
Why can't AI write Arabic correctly in images?
Image models paint pixels that look like letters; they don't type characters. Arabic letters join and change shape by position, small dots change the meaning, and the script runs right to left. Models also appear to have seen far less clean Arabic than Latin text, and several makers say their text works best with Latin letters.
Which AI image generator is best for Arabic text?
As of October 2026, no official page I found promises reliable Arabic. Google and OpenAI say multilingual text has improved, but their own notes warn about spelling and text accuracy. A short word may come out right; check it letter by letter, and plan to set the Arabic yourself.
How do I add Arabic text to an AI-generated image?
Generate the image with no text and a clear empty area, then set the Arabic as live type in Photoshop, Illustrator, Figma or Canva, which all handle right-to-left text. For text on a sign or a package in the scene, add perspective and match the light so it sits in the picture.
Can I make an Arabic logo with AI?
Use AI for ideas, not for the letters. A model can suggest a mood or a composition, but the name must be drawn by hand with correct letterforms, joins and dots. In a logo, the shape of each letter is the brand.
