Text fails because a generator draws shapes, not letters.

A letter is a precise shape with a name, a fixed relationship to its neighbours and a meaning. An image generator reproduces local texture and appearance, not symbols with identities.

At small sizes, letter-shaped texture is enough to look convincing, because the eye reads a word as a single pattern. At reading size the same texture is incoherent, and the illusion collapses.

Longer strings fail more than short ones, because every character must be consistent with the others. A four-letter word can survive; a sentence rarely does.

Typography also carries rules: kerning, baselines, stroke weights, case. A model that has learned what text looks like has not learned what type is.

Each of these appears often enough to be a checklist item, and each is quick to spot once you know it exists.

Product mockups and packaging: a client will read the label, and a wrong character in a brand name is a serious error.

User interface screenshots: generated dashboards with convincing layout and meaningless numbers mislead readers about how a product works.

Signage and environmental text: shop fronts, street signs and posters, where the words are part of the scene.

Documents and print in the frame: a contract, a menu, a book cover, each of which a viewer may try to read.

Illustrations for children or education, where the text is the content and errors teach the wrong thing.

Generate the image with the text area empty, then set the type in a design tool. This is faster than regeneration and produces correct, consistent, editable text.

Brief for the space rather than the words: a blank rectangular label on the bottle, a clean sign board, an empty screen. Describe the surface and the lighting, not the letters.

If a model insists on filling the space, prompt for an abstract mark or a solid colour block, which reads as a logo at small sizes without pretending to be language.

Keep the typeface and the type sizes in the design file so a rebrand updates every asset at once. Text baked into an image is unmaintainable by definition.

Zoom to each text region and read it aloud. The act of reading is the test, and it fails immediately on pseudo-text.

If the language is not one you read, ask someone who does. Pseudo-script is most convincing to people who cannot recognise the real thing, which is exactly the wrong audience to rely on.

Check that text bends with the surface it is printed on. Flat text on a curved bottle is the giveaway in an otherwise convincing product shot.

Then check the hierarchy: real design has weight and size differences. Uniform text is a sign the region was generated as texture.

Replace the region. Cover the generated text with a clean shape that matches the surface, then set real type over it, with the perspective matched.

Or crop it out. Half the time the label was not buying you anything and the image is stronger without it.

If the text is incidental and small, blur or darken it deliberately so it reads as an out-of-focus background rather than an error. Deliberate softness is honest; incoherent detail is not.

Never leave pseudo-text in a commercial asset and hope. Someone will read it, and the reading is the problem.

Generated UI is a special case, because the text is the product. A convincing dashboard with invented labels tells a reader nothing and misleads them about the software.

Build interface imagery from real screens, or from a design file with real strings. It takes longer and it is the only version that survives a close look.

If you need a schematic, use deliberately abstracted elements: grey blocks with three bands, icons without words. Abstracted UI reads as illustrative and cannot be misread as documentation.

For anything a user will compare with the live product, a real screenshot is the only honest option, and it is also the fastest.

Incoherent text is the single most reliable sign of a generated image, which is why it is the second check after hands.

It is not proof on its own. Low-resolution source images, heavy compression and motion blur all destroy legible text, and a real photograph of a distant sign can look exactly like pseudo-text.

The useful inference is one-directional: legible, correctly spelled, correctly set text in a plausible surface is strong evidence of human design, because it is expensive to generate and easy to add.

For assets you publish, that means the text check is also a quality check. Correct text is what separates a generated illustration from a professional one.

If the image carries information in text, that text belongs in the alternative description or in the page, not only in the pixels. Baked text is invisible to a screen reader.

Contrast matters too. Generated text over a busy background is often unreadable for everyone and fails contrast requirements for some readers, which is worth checking before publishing.

Where an image is illustrative and the text is decorative, say so in the alt text rather than reproducing nonsense characters.

These are the same rules that apply to designed images. Generation does not exempt an asset from them.

Some generated typography is close enough to be tempting: a plausible serif headline on a book cover, a sign that almost reads. Close is the most dangerous category, because it survives the glance and fails the read.

If the type is nearly right, decide whether the words are content or set dressing. Set dressing can be softened or cropped until it reads as background. Content has to be rebuilt with real type.

Never patch a single glyph inside a generated word. The stroke weights, spacing and perspective will not match, and the repair is visible at a larger size than the original error.

The same logic applies to logos. A generated approximation of a mark is a brand problem, not a design problem, and it is the one error that cannot be fixed by cropping.

Find every text region, read it, check the character set and the language, then check that it sits on its surface correctly.

If any of it fails, decide: set real type, crop it, or abstract it. All three are faster than regenerating and all three are maintainable.

Then add the text to the alt description if it carries meaning, and publish.

The inspection list on this page runs the pass in the same order every time, which is what makes it reliable across a batch.