Ask an AI image generator for a poster that says “Grand Opening” and you may get something beautiful, with letters that look like a language from another planet. It is one of the most noticeable quirks of AI pictures, and it reveals a lot about how they work. Let us walk through why it happens, and how to work around it.
Why AI pictures put pictures first and words second
Most image generators learn from enormous collections of images paired with descriptions. During training, they pick up patterns: what a cat looks like, how light falls on a face, how a watercolour behaves.
Text in an image is just another pattern, but a demanding one. A human reader treats letters as symbols with strict rules. A model, by contrast, tends to see a sign as a collection of shapes arranged in rows. It learns that signs have letter-like marks, but not necessarily which marks form which words.
Why letters in AI pictures are unforgiving
Think about how forgiving other things are. If a model draws a tree with slightly odd branches, nobody minds. If it draws a hand with a bit of wrongness, we notice, but it is still recognisably a hand.
Letters leave no room. Change one stroke and “O” becomes “Q” or “C”. Swap two characters and a word is wrong. Spelling is binary: right or wrong. A system trained to produce plausible-looking images has no built-in sense that “plausible” is not enough for a headline.
The process makes it worse
Many popular tools that make AI pictures are built on diffusion models: they start from noise and gradually refine it into a picture, guided by the prompt. At each step, the model makes small adjustments. Fine details such as the exact shape of a letter emerge late and at small scale, and they are fragile.
On top of that, the text prompt is converted into numerical representations. Details of spelling can get blurred in the process, especially for unusual words or names. The model gets the gist (“a sign about opening”) rather than the letters.
Why it is improving
The same weakness shows up in chatbots, which stumble on tasks like counting letters in a word; our explainer on why AI chatbots miscount letters covers the text side of the problem.
Newer systems produce better AI pictures. Developers have taken several steps: training on more images that contain clear text; using stronger language understanding to interpret prompts; architectures that handle detail at higher resolution; and separate steps that render text more deliberately.
Short words in plain fonts now come out right fairly often. Long sentences, unusual fonts, curved text and small print remain error-prone.
Practical tips for readable text in AI pictures
If you need text in AI pictures, these habits help: keep it short, because one or two words usually beat a sentence; put the exact text in quotes in your prompt; ask for simple, bold, sans-serif lettering; specify where the text goes, such as centred on a banner; generate several versions and pick the best; and avoid text on curved, tilted or busy surfaces.
If the result is nearly right, fix it manually instead of rerolling forever.
The smarter workflow
If you are choosing a tool for this kind of work, it helps to compare how different services behave. Our comparison of NightCafe and Midjourney is a good place to start.
Honestly, the most reliable method is to separate the jobs. Let the AI create the picture with space left for text, then add the words yourself in a design or photo editor. You get a clean image and perfect spelling, and you control the font, size and layout.
This works for posters, social graphics, product mockups and book covers. It also avoids the awkward situation where a charming image is spoiled by a misspelled slogan.
A warning about trusting AI pictures
When you review a finished image, zoom in on every piece of lettering, not just the headline. Shop signs in the background, labels on bottles and tiny captions are where AI pictures most often slip, and a reader will spot a nonsense word in seconds.
Garbled text is a handy clue that an image was machine-generated, but do not rely on it. As models improve, the tell will vanish. Always consider the source, look for corroborating evidence and be careful with images that claim to document real events.
The reverse also holds: a perfect-looking sign does not prove a photograph is real.
Why this matters beyond posters
The spelling problem is a window into a broader limitation. Generative models are excellent at statistical resemblance and weaker at precise, rule-bound details. You see the same issue with counting objects, drawing correct numbers of fingers, reading clocks and keeping diagrams consistent. Anywhere exactness matters, human checking is still required.
The takeaway
AI pictures are bad at spelling because the models paint shapes that look like writing instead of composing language. Newer models narrow the gap, but short text, careful prompts and a final touch in an editor remain the best approach. Use the AI for what it does brilliantly, atmosphere and imagery, and use your own hands for the words.
