Misspelled or garbled text in an AI ad: TikTok's rule and the fix

TikTok's ad policy has two rules that hit AI video directly: the ad caption and text must be free of spelling or grammatical mistakes, and ad images must not show incomplete or illegible text. Image models and video models often draw words that look right at a glance and read wrong up close. The reliable fix is to keep generated pictures free of text and add the wording afterward. Sume's Video captions burns authored cues or segments you control, with no speech-to-text step.

What does TikTok's policy list?

Under Ad and editorial quality, the page lists the text rule and the visual rule separately. The caption and text rule is the one with spelling examples; the visual-quality rule is written about ad images, and its examples include incomplete or illegible text. Treat both as applying to the frames a reviewer will see.

Why does generated text fail?

A video model composes pixels, not strings. A shop sign, a package label or a UI screen in the clip can come out with a swapped letter, and that is what a reviewer reads as a spelling mistake in the ad. Third-party marks are a second risk: the policy lists blurred or masked third-party watermarks, so a smudge over a stray logo is not a fix.

The safer production pattern is to prompt for a scene with no readable text, then add all wording as an overlay you typed yourself.

How do you burn authored text with Sume?

Video captions accepts cues or segments, each with text, start and end in seconds. Either form skips speech-to-text and burns exactly that copy at those times; the doc calls this the path for silent clips, which otherwise fail with caption_no_speech. Pass one of script_text, words, cues or segments, not several. A standalone job is $0.20 for videos up to 60 seconds, per the pricing note.

The doc also says design overrides do not apply to the punch and tiktok-green styles, so use slam or black-outline if you need to adjust colours.

curl -X POST https://api.sume.com/v1/video-captions \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: overlay-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/example/clean.mp4",
    "style": "slam",
    "cues": [
      {"text": "Fresh from the oven", "start": 0.4, "end": 2.5},
      {"text": "Order today", "start": 2.5, "end": 5.0}
    ]
  }'

Which style should carry the text?

The Video captions doc names slam, punch, tiktok-green, korean-ad and black-outline among its styles. For a short promotional line over a product shot, pick a style whose look is already clear on a small screen, and keep each cue short so the phrase fits in one view. If an omitted style is acceptable, the doc says the wording decides: slam for Latin text and black-outline for Korean.

Because the caption policy objects to excessive or informal use of capitalization, spacing, numbers and symbols among the letters, write cues in normal sentence case with ordinary punctuation. Style effects such as emphasis colours belong to the renderer, not to the characters you type.

How do you check the result?

Pull stills with Video frames at the times your cues show: at[] takes 1 to 24 times in seconds, and png gives lossless images. Read each still on a phone-sized view and compare it with the cue list you typed. If a word is wrong, it is in your list, and fixing it is a new caption job on the clean source rather than a regeneration.

Keep the clean, text-free master. When a caption needs a correction or the market changes, you re-burn on the master instead of stacking text on text. A spell-check pass over the cue list before you submit catches the cheapest errors, including the "Rrom" case the policy itself names.