Every AI video beginner learns this lesson from a shop sign that almost spells a word. The model is not being careless — it literally does not know what a letter is. Once you understand that, the fix is simple and permanent: words are your editor's job, never the generator's.
Look closely at any generated street scene and the illusion cracks at the same place: the writing. Signs trail off into letter-shaped scribbles, book covers wear plausible gibberish, screens glow with keyboards of nonsense. This is not a bug that one more regeneration will cure — it is how these models see the world. This guide explains why, lists what must never be asked of a clip, and shows the division of labour that professional AI-video workflows use: the generator paints the scene; the editor writes the words.
A video model learns from pictures, and in pictures, text is mostly a texture — small dark marks arranged in lines on signs, pages and screens. The model learns that 'a shop sign' looks like a bright rectangle with confident letter-ish shapes on it, because in most footage nobody reads the sign. It has no dictionary and no concept that H-O-T-E-L is a sequence with one correct spelling. So it does what it does with clouds and crowds: it imitates the appearance of text. Short, common words occasionally land by luck. Sentences, prices, names and numbers — the text that carries actual information — fail essentially every time, and a viewer who pauses will see it.
Make it a directing rule: if the viewer must read it, the generator must not make it. That covers shop signs and billboards, book and newspaper pages, phone and computer screens, product labels, logos, number plates, documents and handwriting. Scenes that need these elements can still be generated — a café, a desk, a street — but frame them so the text-bearing surfaces are distant, angled, blurred by motion or out of frame, and end text-heavy prompts with no readable text, no signs, no logo. What a clip may safely carry is everything else: place, weather, light, movement and mood.
The good news: the tool that adds words perfectly already sits on your desk — your video editor. Titles, lower-thirds, labels, captions and callouts added in editing are sharp, spelled correctly, on-brand and translatable. Need a sign to say something specific? Generate the empty sign, then track your real text onto it in the editor, or simpler, cut to a designed title card. This is also how serious channels handle captions for generated B-roll: the clip sets the scene, your text layer tells the truth about it. One workflow, two specialists — the model paints, the editor writes — and the garbling problem disappears from your published videos entirely.
For a tech channel there is a sharper version of this rule. A generated 'computer screen' is not just misspelled — it is invented: fake menus, impossible settings, buttons that do not exist. Showing that in a tutorial teaches wrong information with total confidence. Anything the viewer is meant to follow — a setting, a command, a dialog box — must be a real screen recording of the real software. Generated clips can surround the lesson (the office, the mood, the metaphor) but the instructional core stays genuine. That boundary is what keeps AI assistance from quietly turning into misinformation.
Sometimes a perfect take has one bad sign in the corner. Before regenerating, try the editor's salvage kit: crop the frame tighter to exclude the sign; cover the spot with your caption bar or a graphic; choose the other take — reviewing all variations with an eye for stray text usually finds one clean frame; or let motion save you, since a sign that blurs past in half a second is never read. What not to do is publish and hope: viewers pause, screenshot and share exactly these frames, and one nonsense sign quietly tells them the whole video might be invented.
See the whole process live on screen — every step, exactly as it is done in the video.
▶ Watch Full VideoThey are improving — short words and simple signs land more often than they did a year ago. But 'more often' is not 'reliably', and a sign that is almost right is worse than obvious gibberish because viewers trust it. Verify any visible text frame by frame, and keep information-carrying words in the editor regardless of the tool's claims.
You can ask, and for one or two very common words it occasionally works. Treat any success as luck, not a feature: check the spelling in the actual frames, at full size, before the clip goes anywhere near a timeline. For anything longer than a word or two, do not ask at all — add it in editing.
The same problem, with a legal shadow. Generated logos come out distorted, and a mangled real brand on screen can mislead viewers about who made or endorsed something. Keep logos out of generated footage; if a brand must appear, use the real product shot or your own graphic in the editor.