AI image generators like DALL-E and Midjourney excel at creating photorealistic visuals but struggle with incorporating text accurately. These models don't actually read or understand text; they treat it as patterns of pixels rather than meaningful strings. This leads to bizarre outcomes, like generating a cake with the misspelled phrase “SHOP NUG.” Text is precise, and even a single error can ruin the intended message, while images can tolerate some degree of ambiguity. The process itself compounds these issues. Modern AI uses diffusion models that add and then remove noise to create images. During this denoising, fine details, like individual letters, are deprioritized, which means any errors made early on are hard to fix later. Additionally, the tokenization of text—where phrases get split into smaller parts—can further distort meaning. For instance, “deep focus” may be broken into components that lose the original photographic context. Editing existing images is even trickier. When the AI attempts to fix garbled text, it must maintain the surrounding context while only modifying the highlighted area. This often results in blurry or mismatched text that fails to integrate seamlessly into the image. A recent study highlighted these challenges, revealing that models like Stable Diffusion scored only 1.25 out of 5 for text accuracy in code-related tasks. Across various platforms, the difficulty in accurately rendering structured text remains a significant hurdle.