TikTok ad rejected for undisclosed AI: minor vs significant edits

TikTok's ad policy says undisclosed AI-generated content will be rejected or restricted, and it names minor edits such as lighting adjustments, background removal and denoising as different from significant ones like completely AI-generated images, subjects performing actions they did not do, and AI voice-cloning. Sume's trim, crop and dim run plain ffmpeg with no model; a generated clip is on the significant side.

The categories are quoted from TikTok's page (read 2026-10-10). Where an edit falls in practice is TikTok's judgement, not mine.

The two groups on the page

The page distinguishes substantially altered or generated content from minor enhancement. In the first group it lists completely AI-generated images, subjects performing actions they did not do, and AI voice-cloning. In the second group it lists lighting adjustments, background removal and denoising. The enforcement line is that AI-generated content that has not been disclosed leads to a rejected or restricted ad.

Mapping that to Sume's surfaces

Sume has two kinds of video work. Generation, through the /v1/videos endpoint and its models, creates footage with a model. The media tools are different: video trim, video filter and Timeline use no provider inference and run only worker ffmpeg, according to their docs pages. That distinction is the useful one for a disclosure decision.

The trap in the middle

Assembling a Timeline out of generated clips does not make the result less generated. The render step is plain ffmpeg, but the footage inside came from a model, and the policy language is about the content, not the tool that last touched it. Likewise trimming a generated clip is a cut, not a rewrite; it does not remove the need to disclose.

Voice is where I would be most careful. The page names AI voice-cloning. Sume's voice cloning is available in the app, not through the API, and a stock synthetic voice-over is not named on the page. When in doubt, disclose: the downside of a label is small compared with a rejected ad.

A decision rule you can use

If a model produced any frame or sound in the ad, disclose it on TikTok, using the AIGC label or your own caption. If you only trimmed, cropped or dimmed real footage you shot, check the minor edits list and decide. Keep a record of the Sume steps used for each ad; the result envelope names the kind, such as video_trim or timeline_render, which makes the record cheap.

This table is a planning aid, not a ruling. TikTok's page gives example edits on each side, and real ads fall between the examples. For instance, a crop from 16:9 to 9:16 changes the framing but not the content, and the page lists neither crop nor trim. Rather than argue about where an unlisted edit sits, many teams simply apply the label to every ad that contains any generated frame, and leave labels off for footage that was filmed and only cut.

If an ad is rejected, the page says it will be rejected or restricted, so the fix is a disclosure and a resubmission. Keeping the render settings in your records makes that quick: you can re-run the same trim or Timeline with the same Idempotency-Key discipline, add the caption cues, and upload again.